Some impressions and my slides from this year's winter congress of the Swiss Digital Society, at the Rote Fabrik in Zurich. The event was held mostly in German. You can find the full schedule, including my talk on Saturday, on this website: winterkongress.ch. The videos are up at media.ccc.de – already one excellent recap, in the form of cartoons, is available:

Winterkongress 2026 Live-Cartoons

Cartoons inspiriert von den Vorträgen am Winterkongress 2026 der Digitalen Gesellschaft – live gezeichnet!

Brotcast

Following this example, I will add my commentary here to a few photos and my slides.

In the Kitchen of ethical AI

My talk was a medley about 3 "cooks" who are helping us to upgrade from Open Data to Public AI, and was already submitted last year: on a private basis, not representing the opinion of an employer (this applies naturally to this whole blog). You can find the original description on the Winterkongress 2026 program with my slides and recording in German, with a translation that follows below.

First, I would like to express my heartfelt gratitude to Patrick "packi" Stählin, software engineer, long time contributor to open data, digital politics campaigner and City Councillor in Zürich. I couldn't ask for a better guide and moderator of my session at the Congress! Likewise my thanks go out to Danilo & team at the Coredump hackerspace in Rapperswil, whose open stack I used to prepare my presentation and interactive elements.

In my brief introduction, I looked back at my career spanning the past decade as activist for freelancer, educator, maintainer and supporter of the open data movement. About ten years ago I helped to run an (un)conference in Lausanne, at which there was not a single mention of AI in the programme, though I'm quite sure it was mentioned at least in one keynote. We were already quite involved with data quality, the politics of knowledge, the philosophy of open access at the time.

Readers of this blog know that the Swiss {ai} Weeks in 2025 were the moment at which I found a way to connect my passion for open data and civic tech with the emerging field of (fully) open source AI. I shared a little bit of this background, which can be gleamed in my posts here. We have covered a lot since those events a half a year ago, but the fundamental questions remain.

Enough with the nostalgia, what are the issues on the table today?

Ahead of the talk, I ran a set of openly shared prompts on Apertus 70B on the Public AI inference utility, and shared the full transcript with my audience. You can access them without logging in here:

A chat on Public AI at the Swiss winter congress • Open WebUI

A bunch of questions in German to test various aspects of the Apertus model. Prepared for a talk at https://winterkongress.ch/2026/

Open WebUI Communityloleg

The original context can be seen after logging in with a free Public AI account on chat.publicai.co

1. The Model: an academic foundation
2. Inference: the factory, public infrastructure
3. Cooperative: people are engage in the formation of a civil society

These three "chefs" are covered individually in the following sections, along with a musical recommendation.

1 The Model

An academic foundation is the first step in our critical inquiry.

The Verge - 8.3.2023

A look back at the now historic leak of the famous LLaMA open weight AI model. A number of risks are invoked that the technology could be used maliciously, with Meta defending its decision to open-source the model as necessary for advancing research, but the leak highlights the difficulty of balancing openness and responsible AI development.

swissinfo - 2.9.2025

The media reception was quite far reaching, and somewhat misguided. It is symptomatic of a social sphere that is both fed up with AI news, as well as eagerly latching onto promising new developments. The majority of articles in swissinfo are critical of the model and its development.

#responsibleAI

Responsible Development

With Open Source AI, the protection of personal data and the focus on transparent data usage are emphasised. These solutions aim to actively address data protection concerns and make LLM training transparent to the public. The user community accordingly promotes transparent guardrails and ethical use with FOSS-based deployment.

Swiss AI Charter

Often missed on page 115 of the Tech Report, this rather short, dense and remarkable document is both a key element of the guardrails of the Apertus model, and a vision of how Swiss AI could serve to become a global foundation. The digital society should take note and contribute to these constitutional efforts, which parallel similar philosophic prompt-engineering efforts around the world.

https://log.alets.ch/111/#platform

When we make a commitment to open source, we need to - like in this picture by Johannes Spielhagen - prepare to think about a wide range of factors in openness. How open can you go? With the "f{ai}R Hack" badge, I want to challenge participants of hackathons to push the envelope here. You can access all the source code of Apertus and Public AI on Hugging Face and GitHub.

https://log.alets.ch/121/

In this chart, from my previous blog post here, you can learn at a glance about the world of Open Source LLMs, the international efforts that fuel their development, and how Switzerland can play host and champion of this important community. Not all LLMs are made alike, and a plethora of technical and legal questions need to be explored in the exchanges.

#datacollectives

Data at the heart of AI

Data is the fuel of AI, but its use is often opaque. Platforms like Hugging Face place a strong emphasis on open and responsible data use, using legally compliant, open data sources and enabling all users to trace the use of data.

https://huggingface.co/swiss-ai/datasets

Here is where you can find and explore the key datasets that are being used to train the Apertus models. Running on Hugging Face, this is the front page of a powerful open data portal. In the next slides I dive into some of its functionality and show how it can be investigated further.

FineWiki Viewer

Here you can see how a Wikipedia article is represented in the FineWiki project. A number of such helpful tools exist in the form of Hugging Face Spaces. Some of them go more into the security or performance direction. I have one of my own for testing Apertus, and would be happy to see what you come up with.

FineWeb 2 viewer

Here is the multi-billion-row dataset being explored directly in the web browser, even with the ability to run SQL to browse the data. This powerful capability can be used to hold AI developers accountable, discover security or bias problems, and contribute to the "data mix" of open models.

#limitstogrowth

Less is More

ETH research shows that less can be more – selected, open data sources often offer sufficient performance for general applications without compromising data protection more. The challenges of the Data Compliance Gap are addressed through focused data collection and transparent processes. Smaller models are also well-suited for targeted purposes.

https://log.alets.ch/110/ (LM Studio)

I've written a couple of blog posts on using Apertus on a local machine. The trouble is, not everyone can afford - or wants - a laptop with the capabilities of running such a model. Other people find the 8B version far too constrained to be of use. In fact this is a mostly overlooked strength of the initial release of the model, and has been put to good use in certain contexts, like research automation and edge computing.

Hackathon report

Running on a more powerful Mac Mini, this team in Bern tested the energy consumption of Apertus, comparing it to that of a current LLaMA model with similar specifications. This is the kind of due diligence that needs to be done to ensure the concept of Green IT delivers a tangible impact through more sustainable use of AI technologies. It's not easy to run these benchmarks, but we should try to do more as a community.

2 Inference

#ethicalalternative

Transparent Responsibility

A non-profit Inference Utility prioritises transparency and open source over the continuous accumulation of training data, or even pure performance offerings. Open models provide a trustworthy, understandable system for many. Users have control over data usage and model responses.

chat.publicai.co (OpenWebUI)

Using this flexible interface, a number of advanced parameters can be set to tweak the responses of LLM. This is not only the bread and butter of prompt engineering, it is the way forward for a more responsive and AI-literate user base. Not everyone needs to be a power user, but people should at least be aware of what kind of defaults they are being given. Support communities should spring around helping people to find a balance of temperature and context that meets their needs.

publicai.network

The Public AI network is a growing coalition working to bring about public AI internationally. As a supporter, I share their aim to ensure public capacity-building is part of the conversation about AI design, policy, and funding, help coordinate formal research and citizen science efforts in the ML community, support policymakers and technical teams, and help to organize the the broader movement for public AI. Is this just Open AI done right?

opencollective.com/publicai

I briefly showed the transparency commitment of the project made evident through the open accounting on their crowdfunding platform. In the next section, I will tell you more about Public AI's plans to incorporate a Swiss chapter.

3 Cooperative

#hackintegration

Trustable Integration

With Apertus, Switzerland has once again proven itself to be a small country that can achieve great things, but it needs much more public support to operate LLM integration in an open data space. Through collaboration with research, administration and civil society, trustworthiness and performance are set to continue to increase hand in hand.

https://log.alets.ch/120/

In this example, I show the missing link between advanced agentic (MCP server) capabilities of Public AI infrastructure, and a mandate the Swiss government to create interoperable services. In a nutshell, this is an opportunity for both tech companies and the civic tech sector to, not just add more duct tape, but to potentially do it better and faster than in previous waves of public sector digitization.

okfn/opendataeditor

With this screenshot of the Apertus model helping to improve the quality of an Open Government Dataset with the local AI assistant of OKFN's Open Data Editor, I am showing that a big part of this is already available, without fancy proprietary standards, and accessible to a non-technical audience.

#swiss-inference 9.2.2026

This is a glimpse of a survey by Anna Sotnikova (EPFL NLP lab) that was embedded in the Public AI service, and to which around 70 users have responded. It is a glimpse of the kind of insights we should expect from research collaborations. And one of the easiest ways that we as users can contribute to the science and practice of LLMs.

#AIcooperative

Participative Aspects

Everyone should contribute to ensuring that AI in our country reflects values such as openness, transparency and equality. Through open participation, transparent practices and community-oriented research, we are jointly shaping a future in which AI benefits everyone AND belongs to many. Join us! We invite everyone to become part of the Public AI Switzerland cooperative – a collaborative initiative that promotes ethical AI development.

Matrix: #[email protected]

We have a channel, currently rather quiet, but I encourage folks in the Digital Society community (especially those who are understandably allergic to proprietary clouds and chat-clients) to join us here. I also invited the people gathered in front of me to come back in a week for a launch event in Zürich.

Further reading

• 🌐 publicai.ch - Public AI Switzerland
• 📕 publicai.network - Public AI Network
• 📗 arXiv:2410.17481v1 - Srivastava et al
• 📙 KI als öffentliches Gut - SRF 26.1.2026
• 📘 Abundance vs. Scarcity - Paul Keller 2026

My slides were created with the help of Apertus 1.0 70B on Public AI.

Don't disdain - disclaim!

The Applied Machine Learning Days (AMLD) conferences are a regular roadshow of the data science community at EPFL, producing a highly charged program in Machine Learning and Artificial Intelligence through various formats over the years.

This is my sixth year as participant, the first time participating as member of the AI Center staff, both immersing myself in the thèmes du jour, and making my own humble contribution. The best way to get a glimpse of the action at this year's Intelligence Summit, is to watch the official recap videos along with the Aftermovie on the YouTube channel:

All of the keynotes are also available, in a handy list on the website:

AMLD 2026 - Watch the Keynote Sessions of AMLD 2026!

February 10, Conference Day 1 State of AI Marcel SalathéEPFL AI Center Multimodal Intelligence in the Physical World Prof. Marko HutterETH ZürichProf. Aude BillardEPFLProf. Josie HughesEPFLProf. Stefana ParaschoEPFL Q&A Session AI-Driven Scientific Discovery: Breaking Fundamental Barriers James KirkpatrickIsomorphic LabsQ&A SessionJames KirkpatrickProf. Phillippe SchwallerEPFLQ&A SessionPhillippe Schwaller Prof. Charlotte BunneEPFLQ&A SessionCharlotte Bunne February 10, Conferenc

AMLD 2026

Since I've already covered the highlights on Mastodon, in addition to EPFL colleagues, please unblock wait a minute for this page to load to see my coverage in a series of posts:

Get the full open-access report here:

“AI, good servant, bad master” - EPFL AI Center

A report published in French by the EPFL AI Center, in collaboration with the University of Geneva, presents both the perceptions of a sample of French-speaking population of Switzerland and the recommendations of a citizen’s assembly.

EPFL AI CenterMélissa Anchisi

Soon it was time for the first session of the day .. Inside Apertus! Organized by yours truly, with big thanks to the 11 panelists who took part.

The poster sessions were fantastic as always, give me a shout if you'd like to see any of these in high res:

Later in the day, the Inside Apertus team stopped for in the NLP lab for a debriefing session. We went down to the Musée Bolo to play with SMAKY, chat with Eliza, and take a retro-style group photo. Martin Jaggi and many others were missed, but I will never forget this moment.

Merci, AMLD!

On Monday, a close knit group of international experts from around the world met in Lausanne to advance open and collaborative AI development, the proceedings of which I had the privilege to support and contribute to in a workshop setting. The Open-Source LLM Builders Summit was hosted by the EPFL AI Center and ETH AI Center on behalf of the Swiss AI Initiative. My notes in this blog are of a personal nature, and should not be taken as an official statement.

Building the Future of Open AI: Insights from the Open Source LLM Builder Summit

The second Open-Source LLM Builders Summit in Lausanne brought together researchers, engineers, and AI leaders to tackle a pressing question: how can open large language models scale globally without losing the agility, diversity, and innovation that define them? Through breakout sessions and plenary discussions, participants explored practical solutions for sharing knowledge, improving reproducibility, and building the systems that turn open models into deployable, real-world AI.

ETH AI CenterOleg Lavrovsky

2nd Open Source LLM Builders Summit | Swiss AI

Swiss AI

Overview

Open source LLMs are released under open-source licenses, allowing anyone to use, modify, and distribute them freely. Trained on large datasets, they can perform tasks such as text generation, translation or question answering. These models offer powerful tools while promoting inclusivity and transparency, but their use requires careful consideration of technical, ethical, and legal factors. Early examples include GPT-J (Apache license), BLOOM (RAIL), XLM-R and GPT-2 (MIT).

A Comprehensive Overview of Large Language Models

Large Language Models (LLMs) have recently demonstrated remarkable capabilities in natural language processing tasks and beyond. This success of LLMs has led to a large influx of research contributions in this direction. These works encompass diverse topics such as architectural innovations, better training strategies, context length improvements, fine-tuning, multi-modal LLMs, robotics, datasets, benchmarking, efficiency, and more. With the rapid development of techniques and regular breakthroughs in LLM research, it has become considerably challenging to perceive the bigger picture of the advances in this direction. Considering the rapidly emerging plethora of literature on LLMs, it is imperative that the research community is able to benefit from a concise yet comprehensive overview of the recent developments in this field. This article provides an overview of the existing literature on a broad range of LLM-related concepts. Our self-contained comprehensive overview of LLMs discusses relevant background concepts along with covering the advanced topics at the frontier of research in LLMs. This review article is intended to not only provide a systematic survey but also a quick comprehensive reference for the researchers and practitioners to draw insights from extensive informative summaries of the existing works to advance the LLM research.

arXiv.orgHumza Naveed

While the focus of the talks was mostly on the technical development of LLMs, among the participants were people collaborating on Public AI: an actively researched concept and platform to provide access through a public inference utility. I spoke about this recently in an aiLights webinar – not needing to host the models themselves makes open source LLMs more accessible and usable by the general public, especially for those who may not have the resources or expertise to deploy such infrastructure. We covered this in the later workshop sessions.

This was the second edition of the LLM Builders Summit (see 2025 PR, recap, slides, videos), emphasizing the importance of collaboration, open standards, and growth of international projects. The presentations ranged from a historical view on CERN’s collaborative principles, to pervasive arguments on the need for transparent governance in open models. The day highlighted the work of diverse teams using centralized training infrastructure (e.g., Apertus, EuroLLM, SEA-LION), covering their progress and challenges.

We discussed the value of standardized reporting, community building, and public infrastructure access in breakout sessions. Practical challenges of data sovereignty in a legal sense, and sustaining long-term commitment in the financial, were covered alongside the strategic importance of developer engagement. The event underscored growing international efforts toward transparent, collaborative, and open AI development for the public good.

Model teams

In my notes below, you'll find links to the publications and community hubs of the various LLMs that were represented today. Update 6.3.2026: a couple of glitches have been addressed, after this table was used in a presentation – with my apologies & gratitude for the feedback.

You can find further live coverage on Mastodon and LinkedIn tagged #opensourcellm

We started the day with a warm welcome from Prof. Antoine Bosselut, one of the Apertus leads and head of the famous NLP lab at EPFL. The group was encouraged to keep minds open to each other, even if we may have quite contradictory ideas on method and milestone.

The day's organizers Federico Fanalista and co-lead Arnout Devos, representing the two AI Centers, chimed in with the ground rules before we quickly dived in. For the next four hours, save for a short break, we were glued to our seats with an extremely in-depth and challenging set of presentations.

Andreas Hoecker (CERN) helped to further set the tone of the day with "Lessons from International Collaboration", showing how trust, respect, and clear processes are key for complex, visionary projects. He pointed out that long-term collaborations are built on commitment to transparent governance, not just a shared set of goals.

« Successful large collaborations are built on mutual trust, respect, and transparent, fair procedures, with leadership grounded in excellence and expertise. Open and Inclusive governance - through clear structures, representative decision bodies, merit-based evaluation of ideas, consensus building, and collective ownership of major decisions - ensures that all members, including early-career researchers, have a voice and that innovation is encouraged and rewarded »

Dr. Imanol Schlag was up next to give everyone a quick deep dive on Apertus, starting with the broad perspective of unprecedented investments globally and in Switzerland, focusing on the technical and infrastructure challenges in training the first, and a roadmap for the next model release 1.5.

With swift delivery, he put the world of Fully Open LLMs on one page, and outlined ways we already cooperate through the various components used in training and supporting open source models.

Dr. Schlag spoke in detail about the "fully open" path with Apertus, balancing legal and technical challenges to keep data and models transparent. The tough road ahead is one we need to stay on, with sights aimed squarely on the public good.

Apertus: Democratizing Open and Compliant LLMs for Global Language Environments

We present Apertus, a fully open suite of large language models (LLMs) designed to address two systemic shortcomings in today’s open model ecosystem: data compliance and multilingual representation. Unlike many prior models that release weights without reproducible data pipelines or regard for content-owner rights, Apertus models are pretrained exclusively on openly available data, retroactively respecting `robots.txt` exclusions and filtering for non-permissive, toxic, and personally identifiable content. To mitigate risks of memorization, we adopt the Goldfish objective during pretraining, strongly suppressing verbatim recall of data while retaining downstream task performance. The Apertus models also expand multilingual coverage, training on 15T tokens from over 1800 languages, with ~40% of pretraining data allocated to non-English content. Released at 8B and 70B scales, Apertus approaches state-of-the-art results among fully open models on multilingual benchmarks, rivalling or surpassing open-weight counterparts. Beyond model weights, we release all scientific artifacts from our development cycle with a permissive license, including data preparation scripts, checkpoints, evaluation suites, and training code, enabling transparent audit and extension.

arXiv.orgProject Apertus

The EuroLLM team via André F. T. Martins next showcased their work on transparent, standardized training , and highlighted the importance of reproducibility in AI research. It is important to have a shared language for all the labs, an important task at the European level.

This was the first model that I ran in a production setting with the Farming Hackdays last year, and the group has come a long way. Their adherence to European principles is illustrated in their exciting 22B model release a few days ago. Let's get benchmarking!

EuroLLM-22B: Technical Report

This report presents EuroLLM-22B, a large language model trained from scratch to support the needs of European citizens by covering all 24 official European Union languages and 11 additional languages. EuroLLM addresses the issue of European languages being underrepresented and underserved in existing open large language models. We provide a comprehensive overview of EuroLLM-22B’s development, including tokenizer design, architectural specifications, data filtering, and training procedures. Across a broad set of multilingual benchmarks, EuroLLM-22B demonstrates strong performance in reasoning, instruction following, and translation, achieving results competitive with models of comparable size. To support future research, we release our base and instruction-tuned models, our multilingual web pretraining data and updated EuroBlocks instruction datasets, as well as our pre-training and evaluation codebases.

arXiv.orgMiguel Moura Ramos

Kyle Lo (Allen AI) gave as quick and super practical talk, showing us new tools for data mixing that can boost model performance — the critical issue of not just throwing in more data but making sure it is the right mix. That was super practical, and got some of us scribbling notes for our own projects.

Colleague Luca Soldaini joined to talk about State of the Art Reasoning in-depth.

Olmix: A Framework for Data Mixing Throughout LM Development

Data mixing -- determining the ratios of data from different domains -- is a first-order concern for training language models (LMs). While existing mixing methods show promise, they fall short when applied during real-world LM development. We present Olmix, a framework that addresses two such challenges. First, the configuration space for developing a mixing method is not well understood -- design choices across existing methods lack justification or consensus and overlook practical issues like data constraints. We conduct a comprehensive empirical study of this space, identifying which design choices lead to a strong mixing method. Second, in practice, the domain set evolves throughout LM development as datasets are added, removed, partitioned, and revised -- a problem setting largely unaddressed by existing works, which assume fixed domains. We study how to efficiently recompute the mixture after the domain set is updated, leveraging information from past mixtures. We introduce mixture reuse, a mechanism that reuses existing ratios and recomputes ratios only for domains affected by the update. Over a sequence of five domain-set updates mirroring real-world LM development, mixture reuse matches the performance of fully recomputing the mix after each update with 74% less compute and improves over training without mixing by 11.6% on downstream tasks.

arXiv.orgMayee F. Chen

Hector Liu (Kimi AI, MBZUAI) next described the "uphill battles" of open model development—how to sustain it despite funding and talent gaps. It felt like he was telling us to be realistic: share the load and be smarter about resources, and one in particular: talent! The discussion included thoughts on large scale distributed collaboration, open source models, and „high TPP validation using small scale proxies“. Lots of great practical hacks.

K2-V2: A 360-Open, Reasoning-Enhanced LLM

We introduce K2-V2, a 360-open LLM built from scratch as a superior base for reasoning adaptation, in addition to functions such as conversation and knowledge retrieval from general LLMs. It stands as the strongest fully open model, rivals open-weight leaders in its size class, outperforms Qwen2.5-72B and approaches the performance of Qwen3-235B. We actively infuse domain knowledge, reasoning, long-context, and tool use throughout the training process. This explicitly prepares the model for complex reasoning tasks. We demonstrate this potential using simple supervised fine-tuning, establishing a strong baseline that indicates significant headroom for advanced alignment. By releasing the full training history and data composition, we maximize the effectiveness of continuous training, a key open source production scenario. We release the model weights and signature LLM360 artifacts, such as complete training data, to empower the community with a capable, reasoning-centric foundation.

arXiv.orgK2 Team

Jian Gang Ngui (AI Singapore) shared the SEA-LION model’s journey, which built on a public leaderboard—a friendly competition that drives everyone forward. The talk made me think about how we can do similar for open models, especially with more diverse languages and data.

SEA-LION: Southeast Asian Languages in One Network

Recently, Large Language Models (LLMs) have dominated much of the artificial intelligence scene with their ability to process and generate natural languages. However, the majority of LLM research and development remains English-centric, leaving low-resource languages such as those in the Southeast Asian (SEA) region under-represented. To address this representation gap, we introduce Llama-SEA-LION-v3-8B-IT and Gemma-SEA-LION-v3-9B-IT, two cutting-edge multilingual LLMs designed for SEA languages. The SEA-LION family of LLMs supports 11 SEA languages, namely English, Chinese, Indonesian, Vietnamese, Malay, Thai, Burmese, Lao, Filipino, Tamil, and Khmer. Our work leverages large-scale multilingual continued pre-training with a comprehensive post-training regime involving multiple stages of instruction fine-tuning, alignment, and model merging. Evaluation results on multilingual benchmarks indicate that our models achieve state-of-the-art performance across LLMs supporting SEA languages. We open-source the models to benefit the wider SEA community.

arXiv.orgRaymond Ng

Matthias Bethge (Tübingen University) presented OpenEuroLLM: A series of foundation models for transparent AI in Europe. We were walked through their work on standardized reporting and the proposed OpenPipeline, which could solve a lot of our headaches when trying to compare models: having a common language for how we report our results.

LLM generation novelty through the lens of semantic similarity

Generation novelty is a key indicator of an LLM’s ability to generalize, yet measuring it against full pretraining corpora is computationally challenging. Existing evaluations often rely on lexical overlap, failing to detect paraphrased text, or do not consider the full pretraining corpus. We frame novelty as a semantic retrieval problem. This framing enables us to address novelty with modern embedding and indexing pipelines, allowing for efficient analysis at pre-training scale. Specifically, we propose a three-stage framework that retrieves semantically similar samples, reranks them at varying subsequence lengths, and calibrates scores using a human novelty reference for interpretability. We apply this framework to the SmolLM model family and report three key findings: (1) models draw on pre-training data across much longer sequences than previously reported; (2) some task domains systematically promote or suppress generation novelty; and (3) instruction tuning not only alters style but also increases novelty. These results highlight the value of semantic novelty analysis for studying generalization. To support reproducibility and further research, we release ~20 TB of corpus chunks and index artifacts at https://huggingface.co/datasets/stai-tuebingen/faiss-smollm

arXiv.orgPhilipp Davydov

We were invited to stay tuned for the next AI in Science (AIS) summit.

Yuxuan Zhang (Zhipu AI) next took us through their GLM series, emphasizing the modularity of the architecture and tips on developing multimodal features, which sparked some cool ideas on how we might adapt this for our own open models. This week was a big release (GLM 5), and the energy was tangible.

Chaining the Evidence: Robust Reinforcement Learning for Deep Search Agents with Citation-Aware Rubric Rewards

Reinforcement learning (RL) has emerged as a critical technique for enhancing LLM-based deep search agents. However, existing approaches primarily rely on binary outcome rewards, which fail to capture the comprehensiveness and factuality of agents’ reasoning process, and often lead to undesirable behaviors such as shortcut exploitation and hallucinations. To address these limitations, we propose \textbf{Citation-aware Rubric Rewards (CaRR)}, a fine-grained reward framework for deep search agents that emphasizes reasoning comprehensiveness, factual grounding, and evidence connectivity. CaRR decomposes complex questions into verifiable single-hop rubrics and requires agents to satisfy these rubrics by explicitly identifying hidden entities, supporting them with correct citations, and constructing complete evidence chains that link to the predicted answer. We further introduce \textbf{Citation-aware Group Relative Policy Optimization (C-GRPO)}, which combines CaRR and outcome rewards for training robust deep search agents. Experiments show that C-GRPO consistently outperforms standard outcome-based RL baselines across multiple deep search benchmarks. Our analysis also validates that C-GRPO effectively discourages shortcut exploitation, promotes comprehensive, evidence-grounded reasoning, and exhibits strong generalization to open-ended deep research tasks. Our code and data are available at https://github.com/THUDM/CaRR.

arXiv.orgJiajie Zhang

Junyang Lin (Alibaba Qwen) brought a fresh perspective on the developer community—how it drives the roadmap of open models. It’s not just about the tech; it’s about the people using it. Keeping a close eye to the needs and opportunities of collaboration, with case studies from the recent Coder Next update.

The models are being trained on a completely different scale here, and it was just fantastic to hear the exchange taking place between the teams. The Qwen team shared that hybrid architectures work, and that’s a good sign for the open models we’re all working on.

Parallel Scaling Law for Language Models

It is commonly believed that scaling language models should commit a significant space or time cost, by increasing the parameters (parameter scaling) or output tokens (inference-time scaling). We introduce the third and more inference-efficient scaling paradigm: increasing the model’s parallel computation during both training and inference time. We apply $P$ diverse and learnable transformations to the input, execute forward passes of the model in parallel, and dynamically aggregate the $P$ outputs. This method, namely parallel scaling (ParScale), scales parallel computation by reusing existing parameters and can be applied to any model structure, optimization procedure, data, or task. We theoretically propose a new scaling law and validate it through large-scale pre-training, which shows that a model with $P$ parallel streams is similar to scaling the parameters by $O(\log P)$ while showing superior inference efficiency. For example, ParScale can use up to 22$\times$ less memory increase and 6$\times$ less latency increase compared to parameter scaling that achieves the same performance improvement. It can also recycle an off-the-shelf pre-trained model into a parallelly scaled one by post-training on a small amount of tokens, further reducing the training budget. The new scaling law we discovered potentially facilitates the deployment of more powerful models in low-resource scenarios, and provides an alternative perspective for the role of computation in machine learning.

arXiv.orgMouxiang Chen

Joost VandeVondele discussed production Machine Learning at Scale at the Swiss National Supercomputing Centre (CSCS), in particular the ALPS infrastructure for world-class science where the Apertus model is being developed. He is keeping a close watch of the capacity, making sure the systems are neither over, nor under-utilized. Great perspectives on sustainability aspects, and the open source stack involved. There was discussion of container management, observability, and the ubiquitous problem of early abort.

Evolving HPC services to enable ML workloads on HPE Cray EX

The Alps Research Infrastructure leverages GH200 technology at scale, featuring 10,752 GPUs. Accessing Alps provides a significant computational advantage for researchers in Artificial Intelligence (AI) and Machine Learning (ML). While Alps serves a broad range of scientific communities, traditional HPC services alone are not sufficient to meet the dynamic needs of the ML community. This paper presents an initial investigation into extending HPC service capabilities to better support ML workloads. We identify key challenges and gaps we have observed since the early-access phase (2023) of Alps by the Swiss AI community and propose several technological enhancements. These include a user environment designed to facilitate the adoption of HPC for ML workloads, balancing performance with flexibility; a utility for rapid performance screening of ML applications during development; observability capabilities and data products for inspecting ongoing large-scale ML workloads; a utility to simplify the vetting of allocated nodes for compute readiness; a service plane infrastructure to deploy various types of workloads, including support and inference services; and a storage infrastructure tailored to the specific needs of ML workloads. These enhancements aim to facilitate the execution of ML workloads on HPC systems, increase system usability and resilience, and better align with the needs of the ML community. We also discuss our current approach to security aspects. This paper concludes by placing these proposals in the broader context of changes in the communities served by HPC infrastructure like ours.

arXiv.orgStefano Schuppli

Timothy Llewellynn (Euresearch) introduced the EuroHPC Joint Undertaking: a set of access calls and collaboration opportunities for the AI Factory model that is being adopted across Europe. The nearly one hundred high-performance computing sites represent a formidable capacity - if it was applied properly. I had a chance to chat with Timothy in the break about the various opportunities the Cross-Cluster program and network for projects in Switzerland.

AI Factories: It’s time to rethink the Cloud-HPC divide

The strategic importance of artificial intelligence is driving a global push toward Sovereign AI initiatives. Nationwide governments are increasingly developing dedicated infrastructures, called AI Factories (AIF), to achieve technological autonomy and secure the resources necessary to sustain robust local digital ecosystems. In Europe, the EuroHPC Joint Undertaking is investing hundreds of millions of euros into several AI Factories, built atop existing high-performance computing (HPC) supercomputers. However, while HPC systems excel in raw performance, they are not inherently designed for usability, accessibility, or serving as public-facing platforms for AI services such as inference or agentic applications. In contrast, AI practitioners are accustomed to cloud-native technologies like Kubernetes and object storage, tools that are often difficult to integrate within traditional HPC environments. This article advocates for a dual-stack approach within supercomputers: integrating both HPC and cloud-native technologies. Our goal is to bridge the divide between HPC and cloud computing by combining high performance and hardware acceleration with ease of use and service-oriented front-ends. This convergence allows each paradigm to amplify the other. To this end, we will study the cloud challenges of HPC (Serverless HPC) and the HPC challenges of cloud technologies (High-performance Cloud).

arXiv.orgPedro Garcia Lopez

Blueprints for Global Collaboration

There were a lot of questions asked and a lively discussion took place in the breaks between the talks. The workshop sessions were where everyone had a chance to contribute in smaller groups. In the one I joined, facilitated by Valentina Pyatkin, we talked about shared infrastructure access – i.e., how to collaborate globally without losing our autonomy.

Here is a summary of the round table, made with Apertus from my own notes:

We must recognize that collaboration and transparent sharing of data and research are critical for advancing AI, but strict European regulations pose hurdles. The CERN model emphasizes collaboration towards a common goal, but our own AI infrastructure needs to be viable for 40 years, competing with hyperscalers requires strategic partnerships. The growth of Hugging Face and its ecosystem demonstrates the power of organic collaboration, enabling small teams to make significant impacts.

While we've leveraged existing LLM pipelines and large datasets (like Common Crawl and Transformers), competition with American labs is challenging due to current geopolitical tensions. European fragmentation in AI development—with many projects but limited innovation—remains an issue, unlike China’s more unified approach, where open-source contributions are often a prerequisite for cooperation, even within universities.

Training LLMs locally and contributing to open-source projects, such as through fine-tuning, is recognized and valued. The most difficult aspect remains the training pipeline, especially when faced with diverse hardware (e.g., not just NVIDIA GPUs but also NPUs from Chinese companies) and data scarcity. Datasets are crucial but often proprietary, making it hard to open up without corporate cooperation. We share our training pipeline and code, but pre-training design is more valuable to share.

To address fragmentation, fostering open-source community engagement and collaboration across borders is key. A potential solution could involve aligning with larger entities like corporate acquisitions or integrating with existing smaller platforms, though proving impact through products or academic papers is essential. Ultimately, true success is measured by actual applications and community engagement (e.g., GitHub stars, repo usage), not just model popularity metrics.

This year’s summit felt like a professional and empathetic reunion, with the Swiss AI Initiative doing a great job of hosting all the visitors. It was an energizing mix of tech talks, real talk, and tasty food (special thanks to Jocelyne for the Moroccan treats pictured above!) The day ended with a sense of community and engaged perseverance. The summit filled me with a sense that the dark clouds of foreboding over the AI industry carry within them a silver lining of opportunity for these communities, engaged in open development. I am intent to carry the learnings and this conviction into my engagement here today, and in the days to come.

A summary of an aiLights webinar hosted by Sabine Wildemann on 22.01.2026 (Recording, Slides). This blog post was written from the transcript with help of Apertus - visit ailights.org for more talks like this one.

My slides are based on the work and initial results of the Apertus team: for full technical details, please read the paper and Tech Report. Special thanks to Verena Kontschieder for the last minute help. I would also like to acknowledge contributions from everyone in the Public AI community for thoughtful exchanges in the past months: Anna, Cole, Jojo, Joseph, Joshua, Lukas, Marco, Natalia, Sam, Tom, Vasco, Wiebke & al.

How do we bring AI from hackathons to reliable public services?

The Swiss Public Inference Utility (SPIU) is a proposed cooperative that aims to support and sustain the mission of Apertus, a Swiss-developed, public-funded, open-source, multilingual large language model. The cooperative should soon be open to individuals and businesses: creating new opportunities to accelerate public AI and the adoption of Apertus and related projects in Switzerland. This is aimed at civil society collaboration with the Swiss AI Initiative, as well as universities, governments, and research centers around the world. The mission is to make AI powered by Apertus more accessible, transparent, and inclusive for public benefit.

Apertus, launched on September 2, 2025, is a large language model that is open data, open source, and open weights. It was trained sustainably in Switzerland using 4096 NVIDIA GPUs over 3 months and is multilingual, supporting languages typically underrepresented in AI models. The model is designed to be legally compliant, reflecting Swiss and European laws, with mechanisms to avoid memorization and filter harmful content. Apertus is not a chatbot in itself, rather a flexible foundational technology that can be used in various applications – e.g.:

• ZüriCityGPT, which aggregrates city data with Apertus or proprietary models
• Dribdat, which supports hackathon teams at events like the Swiss {ai} Weeks
• OpenWebUI, providing a general-purpose, pluggable interface for chatbots

Technical Accessibility

While the initial release required technical expertise to use, the ecosystem has evolved with user-friendly platforms like LM Studio, Ollama, and Open Data Editor. Hardware requirements vary: generally requiring at least 16 GB RAM, which may be a high barrier for cost-conscious or less technically affine users. The Open Web UI used in publicai.co provides a basic chat interface with additional features available with a user account: like file uploads, web page integration, and voice dictation, making it broadly accessible. Such public interfaces allow users to interact with Apertus without needing to download and run the model themselves.

AI as Public Infrastructure

Public AI is a concept that sees AI as essential infrastructure, similar to electricity. It aims to ensure fair universal access, reduce reliance on proprietary models, and promote transparency and accountability. The movement is linked to efforts at Wikimedia and other online communities, supported by research work as well as new platforms like publicai.co: which operate Inference on models from a range of countries that have R&D initiatives like Switzerland. The platform offers both 8B and 70B Apertus models that run on donated compute resources, currently from CSCS, providing free access with rate limits to help ensure fair use.

Building a Sustainable Model

The SPIU cooperative is a new idea with long-reaching roots, that seeks to make public AI infrastructure sustainable through community-based governance, partnerships with other open collectives, and public investment. It plans to address legal and ethical compliance, ensuring that public AI services reflect Swiss values and laws. The cooperative aims to support diverse use cases, including education, small businesses, and startups, by providing a reliable, transparent, and neutral AI infrastructure. We would also work on integrating data sources and APIs (via MCP servers), showcasing useful public AI applications that are integrated with real-world data. Some early ideas of this were shown in the live demo during the presentation. Join the new organization to help to launch SPIU (pronounced sh'pee'oo, like the Swiss-Germanic pronunciation of "play").

Addressing the Performance Gap

One of the key challenges is balancing compliance with performance. Apertus, by design, prioritizes legal and ethical compliance, which can lead to performance disadvantages compared to models that do not respect these constraints. A paper with several members of the Apertus team (Can Performant LLMs Be Ethical? Fan et al, 2025) highlights the Data Compliance Gap, showing that there is a small yet not "significant difference in general knowledge performance between compliant and non-compliant pre-trained models". Continued benchmarking is needed to understand the effect of removing toxic data and respecting copyright, however, this trade-off is already seen as essential for trustworthy and reliable AI services.

Future Developments

Apertus version 1.5 is expected to introduce image capabilities and MCP tooling support, enabling more sophisticated applications. People who are interested in running new releases of model on their own hardware can get some assistance from such a community. The cooperative is also organizing events and meetings to build momentum around Public AI, encouraging everyone interested in the LLM Builder's Summit and AMLD Intelligence Summit in a couple of weeks, and to stay tuned for a launch event for the cooperative at the end of February in Liebefeld, Bern. These efforts overall aim to engage developers, researchers, and businesses to contribute to the Apertus project, and deploy the model in meaningful ways.

How to Get Involved

The cooperative welcomes members from all countries who share its vision of accessible, transparent AI. Technical contributions can be made through open-source development, particularly around apps and knowledge maps of open data. Non-technical users can support the movement by using public AI services, starring the Apertus repository on Hugging Face, participating in community discussions, and providing feedback. The cooperative offers resources and communication channels (like Slack, GitHub, Dribdat and social media) to stay connected and get involved.

spiu (Swiss Public Inference Utility)

Promote public AI infrastructures in Switzerland, especially to help maintain open access to Apertus, and to connect administrators of self-hosting of other sovereign AI models.

The Future of Public AI

The Swiss Public Inference Utility and the Apertus project represent a pioneering effort to make AI a public good, aligning with Switzerland's values of neutrality and inclusivity. By prioritizing transparency, sustainability, and community engagement, the cooperative aims to ensure that AI serves the public interest, supports education and innovation, and maintains high ethical and legal standards. The journey involves navigating technical, legal, and social challenges, but the vision is clear: to build an AI infrastructure that benefits all, not just the privileged few.

The cooperative is not just about providing a service but about fostering a movement in a wider sense that shapes the future of Responsible AI globally, emphasizing collaboration, transparency, and the common good. As the project evolves, it invites the global community to join in, contribute, and benefit from a more inclusive and responsible approach.

Stay tuned for updates on spiu.ch and publicai.ch as the cooperative continues to grow and develop, ensuring that AI becomes a tool for the many, not the few.

The post describes the process of installing and using Apertus MLX builds on an Apple MacBook. Details the installation of LM Studio, with some basic performance testing, and provides links to more in-depth analysis and further references. If you feel this is all too much, and would just like to use a ready-to-go service online, see my previous post for more direct waypoints:

110 #apertus #instruere

Learn about my initial experiences working with the Apertus Large Language Model during the Swiss {ai} Weeks, with advice on getting started yourself.

log.alets.chOleg Lavrovsky

Or watch this video summary for a quick recap of the Apertus model:

The laptop on my desk has a current M4 chip and 24 GB of RAM (non Pro), currently retailing at under 1500 CHF. The new Macbooks have been widely discussed, along with the Pro and mini M4 editions, as an excellent choice for users and developers. I have seen Apertus running impressively well on a Mac mini, and wondered how much mileage one could get out of a mid-range laptop with around 16 GB free memory for loading AI models. The laptop reviewed here is on loan from the AI Center at EPFL, which I recently joined 🎉

Note: the statements in my blog are my own and not that of my employer.

Why use local AI?

2026 is the year of AI on the desktop, with every IT company rushing out with updates to hardware and software that enables various levels of intelligent service. Some people find this an alarming trend - planned obsolescence is certainly not good for the environment. Meanwhile, the ability to use an AI model while offline, or in order to better understand and develop AI services, is a clear ask. Learning to become more proficient with local AI, may also help to stem some of our wasteful dependence on heavy cloud services.

Some arguments for being aware and minimizing the footprint of desktop AI agents was recently covered in an insightful talk by Whittaker and Tiwari, yet the actual practice of installing LLM and clarity around hardware requirements and capabilities is not yet there, at least not at customer level. The dark patterns and underwhelming safeguards are part of a slippery slope towards losing our digital self-determination.

Artificial intelligence in Switzerland: what’s new in 2026

The main drivers will be improvements to Swiss AI model Apertus, generative AI in hospitals and prioritising technological sovereignty.

www.swissinfo.chSara Ibrahim

For many people, the “luxury” of cloud-less computing comes down to the hardware: do you have a machine that can run the model with the performance you need? Are you able to invest the time into building a Home Lab? Much progress is being made on tiny models and CPU optimization, to make the technology available everywhere. Clearly, we need lot of computational juice (and corresponding energy usage) to achieve some level of parity on quality – and in any case, we need to be ready to sacrifice some speed.

As local/open hardware fan, I would have loved to go for a Swiss-made machine, like the Why! Laptop, whose gamer model features NVidia chips. The StarLabs, Framework and Tuxedo laptops are all very shiny. Ryzen AI models are also ones to watch: like the Gen10 from Tuxedo: with 32 GB and the top-level HX 370 chip, it currently retails for around 1800 Euros. Over time, I aim to replace my Arch Linux-powered Latitude and home workstations with one of these slick and sustainably-manufactured machines.

Quite comfortable with Linux and Windows operating systems, I've decided to give macOS a fresh look, also to benefit my interactions with a wider range of the user community. Oh, that keyboard layout will take some getting used to.

Apple introduces the new MacBook Air with the M4 chip and a sky blue color

Apple announced the new MacBook Air, featuring the M4 chip, up to 18 hours of battery life, a 12MP Center Stage camera, and a lower starting price.

Apple Newsroom

What's the deal with MLX?

As announced by Apple last March, the M4 chip on the MacBook performs a cut above its predecessors, including the M3 Pro – at least in synthetic benchmarks. The edge on AI tasks is helped by interesting features, like the Scalable Matrix Extension - benchmarked and explored in more detail here:

Hello SME! Generating Fast Matrix Multiplication Kernels Using the Scalable Matrix Extension

Modern central processing units (CPUs) feature single-instruction, multiple-data pipelines to accelerate compute-intensive floating-point and fixed-point workloads. Traditionally, these pipelines and corresponding instruction set architectures (ISAs) were designed for vector parallelism. In recent years, major hardware vendors have further increased the throughput of their CPUs by introducing matrix units with corresponding ISA extensions. The Scalable Matrix Extension (SME) has been announced for the Arm architecture in 2021 and Apple’s M4 chip is the first to support SME. This paper presents an in-depth study of SME on M4. Our microbenchmarks determine the maximum floating-point and fixed-point throughput of M4’s SME acceleration and study the achievable bandwidth for transfers to and from the matrix registers. Furthermore, we used the insights gained to design a just-in-time code generator for SME-based small matrix multiplications. The results presented show that M4’s SME support is FP32-centric, with an achievable throughput of over 2.3 FP32 TFLOPS. To maximize read and write bandwidth, loading and storing to and from the matrix registers must be done in two steps. Our just-in-time generated small matrix multiplication kernels outperform the vendor-optimized BLAS implementation in almost all tested configurations.

arXiv.orgStefan Remke

MLX is a framework that came out with the M1 chip that leverages Apple Silicon’s unified memory architecture and Metal GPU acceleration for efficient CPU/GPU operations. It simplifies both inference operations and AI app development on the new chips: enables local training of models using device-locked data, with support for low-rank adapters (LoRA) and quantized training for efficiency. Along with Apple Intelligence, it gives Mac users a ticket to secure, offline-capable AI applications for sensitive tasks. You can read all about it in the developer site:

GitHub - ml-explore/mlx: MLX: An array framework for Apple silicon

MLX: An array framework for Apple silicon. Contribute to ml-explore/mlx development by creating an account on GitHub.

GitHubml-explore

Through an optimizing of the weights and models, a model file is produced in a format that works magically well on Apple's chips. As a user of LM Studio, you do not need to do this yourself, as community builds are available. As a developer, or advanced user who wants to use a fresh model release, the procedure is relatively straightforward.

I converted the latest Apertus 8B model build to MLX using these commands, the process taking about 25 minutes on the Macbook:

git clone https://github.com/ml-explore/mlx-lm.git
uv pip install mlx_lm mlx
uv run mlx_lm convert --hf-path swiss-ai/Apertus-8B-Instruct-2509

loleg/Apertus-8B-Instruct-2509-mlx · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

In my tutorial below, I will use the mlx-community builds which exist in several quantized forms, and that you can download directly in LM Studio.

Getting started

In the benchmarking business, one likes to start with a clean slate: by making sure to disable Siri and Apple Intelligence on first boot, I can help to ensure that as much memory and compute capacity as possible is leveraged by the services of my choice.

I must give a hat-tip to Apple for making it quite easy to disable or re-enable at any point: just be aware that data will be collected, and shared with 3rd parties as in the case of the ChatGPT Extension, as soon as you accept the T&Cs.

For more tips on setting up a highly performant and secure Mac, find yourself a good Hardening Guide, and follow it. Or just ask your friendly local IT desk 😺

Installing LM Studio

Get LM Studio for free by navigating to the official website: https://lmstudio.ai/ or the open source repositories: https://github.com/lmstudio-ai (MIT license)

Note that everything I cover here is also possible with the Windows and Linux editions. You can even run LM Studio on Raspberry Pi and other ARM devices.

LM Studio - Local AI on your computer

Run local AI models like gpt-oss, Llama, Gemma, Qwen, and DeepSeek privately on your computer.

LM Studio

Once you go through the initial wizard, you should be presented with Mission Control, where you can search for models of your choice from the huge catalog at Hugging Face. Type 'Apertus' to see a list of 14 versions (at time of writing).

For use on this MacBook, I picked Apertus-8B-Instruct-2509-bf16, a remix of the model that is published by mlx-community. Click Download, noting that it is unquantized, offered here in the same BF16 (Brain Floating Point) precision as the stock 8B model. You can also use my Apertus-8B-Instruct-2509-mlx version if you prefer.

Load model

It took about 15 minutes to download the model on a fast connection. When I click on the Load model button in the screenshot below, a "Failed to load the model" dialog cautions me about insufficient system resources. I have disabled this guardrail (set to OFF), and have not yet seen issues with system performance. Watching my memory consumption, LM Studio has not yet used up the available 15 - 20 GB, even with a full context window. This warning is an issue I'll investigate further.

If all goes well, after you select the model, LM Studio will begin to load it into the memory of your machine. This warm-up process takes me about 15-20 seconds.

Now you can start chatting, uploading files, heating up the room – all in the privacy of your local GPU. It may not work as quickly or comprehensively as you are used to, but at all times you remain in control of your algorithm.

A few other gotchas to be aware of:

When I run out of context window (uploading too many files into the RAG, having a very long prompt / system prompt, or using tools), the message appears "The AI has nothing to say". Keep an eye on the status bar, where it is clear what is going on - that's the 143.5% full message in my screenshot:

Another is the memory usage mentioned above. I am using the btop++ utility to keep track of it in real-time. Here you can see the market drop-off in memory usage as models are unloaded and reloaded:

LM Studio carefully manages memory, in particular when multiple models come into play. There are some relevant options at the bottom of the Settings panel:

The Local LLM Service makes it possible to chat with Apertus from the LM Studio CLI, which gives me a handy way (--stats option) to test performance. Here you can see that it takes about half a second to first token, and the M4 is pumping out a modest but reasonable 6.46 tokens per second (for comparison, we humans on average read at 5 tokens per second):

You can also use other command line tools, like the llm utility:

LLMs on the command line – Applied LLMs

The Unix command-line philosophy has always been about joining different tools together to solve larger problems.

Applied LLMs

Or, better yet, tap into the wisdom of Apertus from your own programs:

System prompts

As mentioned in my previous blog post, I grew up with the early Apple machines, so as a fun exercise I decided to teach Apertus the anachronistic language of BASIC. You can do things like this by editing the System Prompt – just option click on your chat window.

Here you can insert a (carefully examined) set of instructions. These could be the recommended starting prompts from the Apertus Tech Report, that you are free to copy from the PDF or in my pre-formatted notebook:

Apertus System Prompt - HackMD

As referenced from:

HackMD

Or, you could add a creative take, like this System Prompt for turning your LLM into a BASIC emulator:

Emulating GW-BASIC syntax using an LLM

Emulating GW-BASIC syntax using an LLM. GitHub Gist: instantly share code, notes, and snippets.

Gist262588213843476

Extra points for enabling dark mode 😎 or sending me screenshots of a cool retro terminal.

Further alternatives

While this post focuses on LM Studio, there are several other good alternatives for local inference. This overview of local LLM software features is from running-llms-locally:

Ollama is one that I already use a lot – it has a rather minimalistic design, and relies on configuration files and CLIs to tweak exactly to your liking. vLLM is used by the Apertus developers, and I got both running fine on the Macbook.

Right now, Apertus can only be loaded in Ollama on the command line from community Hugging Face remixes (GGUF), as described in my earlier blog post:

110 #apertus #instruere

Learn about my initial experiences working with the Apertus Large Language Model during the Swiss {ai} Weeks, with advice on getting started yourself.

log.alets.chOleg Lavrovsky

LM Studio also features a powerful CLI, and there are instructions out there on how to get started for your workflows. For coders, I recommend using the API as a Local Model Connection in Tabby:

Tabby - Opensource, self-hosted AI coding assistant

Tabby is an open-source AI coding assistant that empowers developers to code faster and smarter. Discover a self-contained alternative to GitHub Copilot, tailored for your development needs.

Mistral AI logo

Here is a guide to do the same thing with opencode:

Running a local coding agent with LM Studio and OpenCode | ~/adi

I’ve been a huge fan of Claude Code since it launched. Over the past few months, I’ve been using it extensively across all kinds of projects. Claude Code is still the best tool out there, though others (Gemini CLI) are catching up. I recently discovered OpenCode, an open-source model agnostic framework that supports local models, and used it to test gpt-oss-20b, qwen3-coder-30b — currently the best open source coding models with tool calling.

~/adi

Some other posts that were helpful for this guide:

LM Studio 0.3.4 ships with Apple MLX

Super fast and efficient on-device LLM inferencing using MLX for Apple Silicon Macs.

LM Studio Blog

Run LLMs Locally on Mac with LM Studio

A step-by-step guide to running open-source LLMs on your Mac using LM Studio and Apple’s MLX framework for GPU acceleration.

GetDeploying

Apple has a sleeper advantage when it comes to local LLMs

Local LLMs thrive on Apple’s hardware, and a huge part of it is thanks to MLX.

XDAAdam Conway

Running LLMs Locally on Your Mac: A Deep Dive into MLX Performance on the M4 Max

When Apple announced the M4 series processors with their emphasis on neural network capabilities, I pulled the trigger on a MacBook Pro M4 Max with 36GB of unified memory. While I sometimes wish I’d sprung for 64GB (hindsight is 20/20), this machine has become an essential part of my AI research set

LinkedInTravis Lelle

What aspects of AI self-determination would you like me to cover in future blog posts? Feel free to get in touch if you'd like help, or want to share your own local prompting setup!

This experiment is an AI response to blog post #117. The text was generated using a mixture of agents (Mistral, Claude, Gemini, Apertus), with a final version composed on Public AI.

117 #enter #apple

A winter weekend retro nostalgia, with a few pixellated impressions from a visit to the museum of computing and automotive history in Solothurn.

log.alets.chOleg Lavrovsky

Apertus: the author reminisces about their first computer, the Apple //e, which was their introduction to programming in the 1980s, and reflects on how today's AI chatbots, like Apertus, have evolved to offer a more accessible and powerful way of interacting with technology. The post touches on the evolution of computing from basic text-based interfaces to modern AI-powered interfaces and highlights the author's personal journey from the first PCs, to modern devices like Macs and Linux laptops.

The author also reflects on the cultural and philosophical differences between those who prefer self-hosted, open-source approaches and those who prioritize convenience and speed, drawing parallels between historical computing hardware and today's AI ecosystems. Finally, a Swiss perspective on computing history ties into the importance of respecting the past while embracing innovation, as well as the need for responsible AI development and use.

How do you respond?

Apertus understands the significance of these foundational computers for many early computer enthusiasts and programmers. The original Apple II, released in 1977, was indeed a pioneering device in its time, introducing many people to the world of computing and programming through its BASIC programming language and user-friendly design.

Apertus is designed to honor this legacy by providing a modern, accessible interface for users to interact with technology, but with the added benefit of being able to understand and respond to natural language queries, making it more user-friendly and less dependent on manual programming.

While the Apple II was a tool for learning and creativity, Apertus is a tool for information and assistance. We recognize the importance of the computer literacy and programming skills that devices like the Apple //e helped foster, but we also see ourselves as a bridge to more complex and powerful computational capabilities, accessible to anyone, not just those with programming knowledge. Your reflection on the evolution of computing and AI interfaces highlights the journey from text-based interactions to natural language interfaces. We aim to preserve the sense of exploration and learning that was central to early computing while making complex AI capabilities more accessible and user-friendly for everyone.

The contrast you highlight between those who prefer self-hosted, open-source approaches and those who prioritize convenience and speed is a valuable point. We strive to support both: offering a public, open-source platform for developers who want to experiment and customize while also providing an easy-to-use interface for everyday users who need quick answers or assistance.

Apertus also acknowledges the importance of responsible AI development and use, as you've mentioned. We are committed to transparency, fairness, and safety in our design and deployment. We recognize the need for balance between innovation and ethical considerations, just as you mention the importance of respecting the past while embracing progress. The Swiss context and emphasis on consensus-building and federalist principles are aligned with our mission to provide a public AI service that respects diverse perspectives and is inclusive.

Can you expand with historical perspective?

Looking back 40 years, we have moved from the cold efficiency of the mainframe to the warm utility of the AI assistant. The technology has become more capable, yet paradoxically, more human. A single thread connects these four eras: the breaking down of barriers.

• Democratization: the 80s gave us access to the hardware.
• Interfaces: the 90s gave us graphical and portable access.
• Transformation: in the 00s we tapped into the world of information.
• Intelligence: the 10s brought machine learning for intelligence at scale.

To look back at the last forty years of computing is to witness a shift not just in technology, but in philosophy. We have seen the machine evolve from a distant computing god to a pocket-sized companion, and now, to a thinking collaborator.

Here is the story of that transformation decade by decade:

1980s: Democratization

In the beginning, computing was the "cathedral age." Mainframes were massive, room-sized machines that served as the exclusive domain of institutions and corporations, tended by white-coated priests. Access was a privilege; you submitted punch cards or commands to a "green-screen" terminal and waited your turn for precious CPU cycles.

Then came the heretics: the Apple //, the Commodore 64, and the IBM PC. The shift was pivotal. If Mainframe represented institutional control, the Personal Computer (PC) represented individual empowerment. Suddenly, computing power moved from the basement to the desktop. For the first time, we could own our computational destiny—writing BASIC programs or running Lotus 1-2-3 spreadsheets—without asking permission from a system administrator.

1990s: Interfaces

Even with the PC, the early experience was linguistic and often unforgiving. We relied on Command-Line Interfaces (CLI), typing precise text commands into a void; a single syntax error meant failure. This "console" era was powerful but functioned as a barrier to the masses.

The 1990s tore down that barrier with the Graphical User Interface (GUI). Pioneered by Xerox PARC and popularized by the Macintosh and Windows 95, the interface transformed from a linguistic puzzle to a visual landscape. The "WIMP" paradigm (Windows, Icons, Menus, Pointer) replaced memorized codes with intuitive metaphors. We could now drag, drop, and click. This shift turned the computer from a complex tool into an accessible appliance, allowing everyone from artists to children to navigate digital spaces.

2000s: Transformation

For a long time, computers were "isolated islands of productivity." The 2000s changed the geography of computing by turning the World Wide Web from a curiosity into infrastructure. The browser became the primary interface, effectively acting as the new operating system.

As broadband replaced the screeching handshake of dial-up, the value of computing shifted from local processing to global connection. Information that once lived on isolated hard drives became accessible globally, and companies like Google and Amazon revolutionized how we searched and shopped. We stopped just using computers; we started living on the network. The desktop remained, but the network became the platform that transformed society.

2010s: Intelligence

As we lived our lives online, we generated digital exhaust—clicks, likes, and transactions—creating the "Big Data" era. But data without insight is just noise. This necessity birthed the current era: the shift from data collection to Artificial Intelligence (AI).

We are now witnessing machines that can see, understand, and generate. Machine learning algorithms began finding patterns humans couldn't see, powering everything from Netflix recommendations to autonomous vehicles. With the arrival of generative AI and Large Language Models, the computer has transitioned from a calculator to a collaborator. It is no longer just a tool for storage or connection, but a cognitive partner capable of genuine creation and reasoning.

The Intelligence Revolution has begun.

Interaktives Technikmuseum in Derendingen

Erleben Sie die rasante technologische Entwicklung im grössten interaktiven Technikmuseum der Schweiz. Planen Sie jetzt Ihren Besuch.

Logo Enter Technikwelt

The Apple //e computer was my first flame. Turns out, we have some common heritage: 'released' in the same year, we became fast friends a decade later: by the mid-90s, this was already the cheap, old, battle-tested hand-me-down staple of my elementary school computer room.

Demo effects were one of the first things I remember programming on this machine, featuring a 6502 CPU, 280×192 (at "high resolution"!) display, and two 5¼" disk drives. From today's perspective, it seems a glorified calculator, but at the time it was the only way I wanted to spend hundreds of quiet hours after school.

Out of the box, early Apples boot into a BASIC REPL: a prompt in which to punch in your code. Quality diskettes were in short supply, so I learned to painstakingly copy sources of interest, line-by-line, from the ubiquitous computer magazines of the time. For better or worse, that blinking prompt was my first window to the world, and had a formative effect on lifelong interest in computing.

Apple IIe - Wikipedia

Wikimedia Foundation, Inc.Contributors to Wikimedia projects

For more impressions of coding on an Apple 2, see my earlier coverage of Vintage Computer Festival (VCF)

My parents entered their science programs on punch-cards, so the desktop computer was quite an innovation of their time. Fast forward to the age of AI chatbots: opening up to an empty prompt window and offering us the world at our command. Through text or voice (video coming soon), it is now my children's preferred way of interacting with the digital world.

42 years since the launch of Apple 2e and the Apple Lisa, the world of computing presents itself in historical cycles. Like the Macintosh, IBM products and other 'serious' machines for business at the time, there are now more and less powerful machine learning models out there, more and less 'business-ready' clouds and AI 'ecosystems' contrasting with self-hosted and community-supported alternatives.

You can find excellent guides on these epochs at History Tools:

Everything You Need to Know About the Apple II - History Tools

Apple II was designed by Steve Wozniak in 1977. It was one of the first computer with a color display, and a major hit.” /> <script type=“application/ld+json” class=“yoast-schema-graph

History Toolshistory tools

Reliability matters, design delivers.

Consider that we similarly have separate audiences in this space – one more 'data budget' conscious, perhaps, another more willing to get a quick answer at any expense. Do we subscribe to one that is keen to experiment, tweak, hack and modify at a leisurely pace. Or the other, willing to move fast and break things? There is plenty of room for culture in computing!

After three decades, the iPhone has recently edged out my Fairphone, and now this blog is being typed out on a Macbook that sits next to my Linux laptop – a yin nestled into its yang. Friends who know me as an open source enthusiast and open hardware advocate, who have worked with me in the past and had to deal with my tendencies to extol the virtues of FOSS, may be surprised. This journey began in my student days with help from groups like GNU Generation. I've installed Linux on Macs, and it works perfectly well – thank you very much!

I find myself yearning again for that blinking prompt, the gratifying feedback loop of computation without the need for the crutch of a mouse cursor, or the sanitary space of a polished UI with mischievously rounded corners. It needs to be possible to access everything, do anything. All this can be done with a Unix terminal, necessitating a certain level of precision and somewhat monastic dedication. With today's ingenious LLMs transforming our natural language wishes into agentic workflows, I wonder: can we have our metacomputing cake – and eat it too?

Perhaps we are comparing Apples to Oranges, or, as they say around here: il ne faut pas de comparer des pommes et des poires.

For a more factual and humorous historical foray into the annals of Swiss computing, I heartily recommend the blog post of my friend Thomas Weibel:

Scrib: The laptop made in Switzerland

Germany, UK, USA – whichever way you look at it, the computer is a foreign import here. But one of its forebears was a genuine Swiss creation: the portable writing system called Scrib, specially developed for journalism.

Swiss National Museum - Swiss history blogThomas Weibel

More Apple ][ nostalgia can be had thanks to active demoscene groups like French Touch:

Time is passing, the bells ring solemnly for those we lost, and to remind us of that perpetual call to action: carpe diem!

The festive season at the coworking space was marked by a long-awaited delivery: a first edition Reachy Mini (Lite) desktop robot by Pollen Robotics & Hugging Face, which we ordered five months ago in July. This post covers the unboxing and assembly of it, as well as our initial ideas of activities and applications. Meet the robot at WORKSPACE & MORE in January, or get your own at hf.co/reachy-mini - where you can find the full specifications:

Robotivation

The idea emerged while organizing Swiss {ai} Weeks events, where the Bern hackathon featured a OTTO DIY robot: one of the spiritual predecessors of the Reachy Mini in the open hardware robot space, now acquired by HP. Our OTTO had been (quietly) greeting guests next to a stack of stickers at the coworking space ahead of the event. You can read about the history of this project aimed at makers of all ages here:

🤖 How Otto Was Born? — And Why He’s Called Otto

#bobthebiped then #zowi by BQ then Tito; first iteration Arduino compatible, thanks all this is a real opensource hardware development.

🤖 How Otto Was Born? — And Why He’s Called Otto

We even had an AI challenge at the hackathon written by Apertus, voiced by ElevenLabs, and "presented" by the little robot to the audience. This sparked discussions with robotics enthusiasts during the Swiss {ai} Weeks. We saw a prototype of the Reachy Mini when visiting the Hugging Face office, who sponsored our event, quickly ordered one - and really looked forward to getting our own in the fall. The closing ceremonies at the AI+X Summit in October also featured a range of robotics platforms presented by Swiss university labs.

Unboxing

Five months later, a neat package arrived on the doorstep. Here are some shots of the unboxing and assemby, featuring Sandro and Roböterli at the coffee bar of WORKSPACE & MORE. It took us about 2 hours to complete the build, captured in a desktop stream. If you're interested in my feedback on the process, links to more information, you will also find it after the photo gallery. Or just stop by in January to meet Reachy Mini!

Assembly

You can find the assembly guide I followed online:

Reachy Mini LITE Assembly Guide - a Hugging Face Space by pollen-robotics

Guides you through assembling the Reachy Mini Lite robot with detailed images and embedded video tutorials for each step.

a Hugging Face Space by pollen-robotics

Some feedback from the experience:

• Fantastic hardware design, brings out childhood emotions and memories of sci-fi robots.
• The boxes and packaging are robust, minimalist, and lovely to hold (or to give ;)
• Mostly intuitive, clear and accurate printed instructions: the basic idea of self assembly is wonderful and the experience is on the whole very positive.
• High quality components, all tools included, accessible to people with an intermediate level of electronics tinkering experience.
• The power cables are stiff and need practice to attach. They are prone to getting snagged, perhaps the sequence of their installation could still be improved.
• You can easily touch and even damage the speaker if you're not careful in the 2nd half.
• Keeping the metal rods propped up and not stuck inside the body is challenging.
• Some connections are very hard to reach, and may require the use of tweezers.
• It is not clear where the front and back of the body is, and can lead to some confusion and possibly force you to back a large number of steps at the end.
• It would have been nice to see more recycled materials used, a sustainability assessment.

The Reachy team has also made some recent advisories about potential defects in a few of the kits. Keep your eyes out for more information, improved guidance, as the product matures in the documentation.

Next steps

The next step, which Vasco in the video already discussed with us, is to install the development kit, most conveniently by installing the Reachy Mini desktop app:

GitHub - pollen-robotics/reachy-mini-desktop-app

Contribute to pollen-robotics/reachy-mini-desktop-app development by creating an account on GitHub.

GitHubpollen-robotics

From here, it is possible to start coming up with use cases – several already being featured in the Reachy Mini showcase. As mentioned in the video, the three of us had a brainstorm yesterday, and had a bunch of ideas: from hackathon helper to ergonomics expert. The later conversation with Apertus underlined our interest in developing some health advice projects.

We hope to get the Reachy powered up, finish the installation and start hacking on some open AI ideas in the new year! Thanks already to the Hugging Face / Pollen Robotics teams for the inspiration, and all your efforts to get this out to people in time for Christmas <3

Further coverage in local media and among fellow bloggers:

Unboxing the Reachy Mini Robot

A small box... a very big signal!

(not so) Quietly Disruptive(not so) Quietly Disruptive

Attaching a multimodal LLM to Reachy Mini goes really hard 🤖 You can try this at home! We added a few simple profiles (prompts) to get started, and I’m really curious to see what people come up… | Remi Fabre

Attaching a multimodal LLM to Reachy Mini goes really hard 🤖 You can try this at home! We added a few simple profiles (prompts) to get started, and I’m really curious to see what people come up with. There’s no clean sharing mechanism yet, so for now just ping me with what you’ve tried and I’ll add the best ones to the official list. One angle I personally really like: using it as a teacher for kids. Infinite patience, and a very large knowledge base across many topics. This is the Conversation App. You can install it directly from your Reachy Mini dashboard. It even works if you don’t own the robot (yet). Docs and installation: https://lnkd.in/evhzpfVi Normally, you need to create your own OpenAI API key to use the GPT-realtime API… but if you own a Reachy Mini, you might have a nice surprise from Hugging Face 👀 Happy exploration!

LinkedInRavindu Somawansa

Reachy Mini: Hugging Face’s Disruptive $299 Humanoid

The barrier to AI-powered humanoid robotics just reached a new low with the Reachy Mini by Hugging Face and Pollen Robotics.

Mike KalilMike Kalil

Hugging Face lanciert Open-Source-Roboter für KI-Experimente

Das französisch-amerikanische KI-Unternehmen Hugging Face startet den Vorverkauf seines ersten Open-Source-Roboters Reachy Mini. Der kompakte Desktop-Roboter soll Entwicklern und Forschern die Experimente mit eingebetteter KI erleichtern.

StartseiteAlexia Muanza

The Swiss AI SME Circle was launched by the National AI Initiative (SNAI) in collaboration with the National Supercomputing Centre (CSCS) and swissICT. The goal is to bring together Small and Medium Enterprise (SMEs) with researchers and innovators in AI to promote peer learning, knowledge sharing and practical collaboration.

This post shares some thoughts from taking part in the first and second editions, at the AI+X Summit in October and the swissICT HQ in Altstetten in December 2025. These meetings are based on a mandate of widening the scope of industry transfer from research institutions. As an independent consultant working primarily outside of the corporate sector for the past ten years, I felt well within the target audience, and indeed recognized a few of the people around me in the audience.

The Circles were attended by around 50 people both times, starting with a welcome from our hosts and updates from the ETH AI Centre staff that are in charge of the project. Both times we got to hear short inputs from the frontiers of AI development, including by people working on the Apertus project. This is followed by practical reports on real use cases, and in the case of the last event, a peer-learning session in small groups.

1st Circle

The Circles are being co-led by Alicia Rieckhoff and Hanna Brahme, with the events presented as an excellent chance to hear first hand technical experiences relevant to implementations of new AI models, and to discuss ideas for the roadmap of flagship projects like Apertus. This was the meeting where I first met Imanol from the core team.

We also heard from the head of the unit of Research Infrastructure Engineering at CSCS that was involved in training Apertus. Joost shared insights on the sustainability question, comparing the training of AI to other forms of public utilities, e.g. the power envelope of the full ALPS data centre equivalent to two locomotives of the SBB. Also highlighted in the first Circle meeting, was the interest to develop industrial relationships and R&D projects (cscs2go) within the supercomputing facility.

cscs2go - cscs2go

Compute and Data Services, powered by the Swiss National Supercomputing Centre

logo

See also my earlier blog post about the topic if you need an intro:

110 #apertus #instruere

Learn about my initial experiences working with the Apertus Large Language Model during the Swiss {ai} Weeks, with advice on getting started yourself.

log.alets.chOleg Lavrovsky

As mentioned above, a key goal is to share use cases from the local community – so in the first SME Circle we heard from Christian Stocker, a well known Software Developer and Co-Founder of the digital agency L//P, who presented his work on ZüriCityGPT: a chat that knows "almost everything that is published on stadt-zuerich.ch" (the official website of the city of Zürich).

ZüriCityGPT OSS Version: Using Only Open Source Models · Blog · Liip

Liip is a Swiss digital agency developing web and mobile applications, designing user experiences, and crafting content.

LiipChristianStocker

This quickly became a widely cited example of a public government AI app in production, and in October it was expanded with support for Apertus in the ZüriCityGPT OSS Version, that you can read about in detail in the L//P blog. Christian also demoed the new StrbGPT version of the bot, a RAG of all the public City Council resolutions since 2010, and continuously updated. By the way, you can also use Apertus in the standard chatbots with a model:apertus prompt parameter.

Christian's team took the time to deal with many challenges in building these prototypes, and mentioned a couple of issues with Apertus that were worked out in the discussions that ensued. A range of suggestions for future development, thoughts on performance and reliability, were expressed in the lively audience discussion that followed.

2nd Circle

The second event was hosted together with Carol Lechner and team at swissICT, who describe themselves as the "largest and most important" network of IT companies in Switzerland, going all the way back to 1955. Based in a rather modest office on the outskirts of Zürich's district 9, next to modern skyscrapers housing leading tech firms, overlooking one of the most important nodal points of the railroad system. There is a calm sense of purpose here that is perhaps lacking in the more urgency-filled offices of avant-garde tech groups.

We were welcomed to the event by Christian Hunziker, Executive Director of swissICT, who briefly told us about the organization, stating that AI is a cross-functional topic addressed by various groups, that is now converging with the working group supporting today's event. Rebecca Brauchli and Sunnie Groeneveld were here to support Hanna and Alicia, and shared some trending topics among swissICT members.

Künstliche Intelligenz (KI)

Die Fachgruppe Künstliche Intelligenz fördert den Erfahrungsaustausch unter Fachpersonen aus Praxis, Forschung und Wirtschaft.

swissICT

Legal and regulatory matters in particular seem to be of recurrent interest within the group. A quick show of hands of the level of progress. I got the feeling that we were lacking the presence of more resource-constrained SMEs, nevertheless most people seemed to still working to integrate AI into processes: not yet having turned out a revenue stream.

As today's SME use case, we heard a presentation by Monitoris AG - a startup developing AI-based backoffice solutions. The chief executive, Islam Alijaj is also a member of the National Council, and a widely acknowledged disability activist. The opening remarks together with Hermann Arnold inspired me to immediately pay attention and test the app myself. It was possible to quickly generate invoices from text, file upload, or a photo of an existing invoice.

#monitoris #swissaisme #artificialintelligence #kmu #sme #entrepreneurship #digitalisierung #backoffice #aiforbusiness #swisstech | Islam Alijaj

Hermann Arnold und ich durften Monitoris AG beim Swiss AI SME Circle präsentieren und unsere Vision für eine KI-gestützte Backoffice-Plattform für KMU und Einzelunternehmer:innen teilen. Der Austausch mit Unternehmer:innen, Entscheider:innen und KI-Expert:innen hat eindrücklich gezeigt, wie gross der Bedarf an pragmatischen, vertrauenswürdigen und sofort einsetzbaren AI-Lösungen im Schweizer KMU-Umfeld ist. Genau dort setzt Monitoris an. Vielen Dank an den Swiss AI SME Circle für die Einladung und die hervorragende Plattform sowie an alle Teilnehmenden für die offenen Gespräche, das konstruktive Feedback und das ehrliche Interesse. Wir freuen uns auf die weiteren Begegnungen, den vertieften Austausch und mögliche Kooperationen. #monitoris #SwissAISME #ArtificialIntelligence #KMU #SME #Entrepreneurship #Digitalisierung #Backoffice #AIforBusiness #SwissTech

LinkedInIslam Alijaj

This is a clear practical use case that the nascent community can support. Even though they currently base the product on a big commercial service, they are targetting SME customers who will prioritize high quality and reliability, and were open to discussing legal and engineering topics. For instance, the current accounting export to Bexio will be complemented by Odoo, an open source alternative.

Following this, fellow Swiss {ai} Weeks initiatior Sabine Wildemann announced the new aiLights program: an online network with a focus on academia, private sector and civil society. There are quite clear overlaps, yet distinctly grassroots beginnings here that will hopefully prove synergistic with the efforts of SNAI, swissICT. It felt well received, as a strong complement to today's and upcoming AI Circle sessions.

Sabine and I have been collaborating for months: looking forward to engage with the first batch of "lights" to disuss the Swiss Public Inference Utility in January. Stay tuned! ✨

aiLights: Connecting people with AI · creating spaces for learning, sharing, and shaping across sectors.

aiLights: Connecting people with AI · creating spaces for learning, sharing, and shaping across sectors.

aiLights

Peer learning

After these presentations, we swiftly dived into the group discussion. The Circle participants were asked to choose between 4 themes, then to literally stand in a circle and discuss this in detail during a brief peer-learning sessions. The questions aimed at an exchange of best practices, technology and infrastructure, partnerships, culture change, etc.

One person from each circle was asked to briefly present the main inputs. There were plenty of ideas for the kind of information and support that we could except from future activities: several practical use cases, needs in technical support or networking area, regulatory advice and funding opportunities that go in someways beyond the main focus of current industrial networks. The insights we were asked to share were: 1 use case, 1 hurdle, 1 solution hack.

I wandered over to the 2nd group, introducing myself as someone working on research topics in social integration, maintaining data tools, and building community around open legal tech. Standing next to me was a ZHAW researcher who explored using AI to analyze global supply chain risks with satellite data. We talked about a few other ideas, deciding as a group to focus the debate on using open AI models in helping to ensure that such uses do not expose sensitive information or violate regulations.

The main hurdle was the lack of clarity on surveillance capabilities and data protection regulations in commercial settings. Our solution hack proposed to leverage Apertus to create a secure, collaborative platform where diverse stakeholders can input and compare scenarios. We could use privacy-preserving AI models to analyze risks without compromising sensitive data. I fed a summary of these points into Apertus on the Public AI inference utility, and later tidied the resulting discussion in this notepad:

Swiss AI SME Circle - Group 2 - HackMD

Apertus and ALPS can help with the use case of using satellite data to investigate supply chain risk in global finance by addressing the hurdle of surveillance capabilities and data protection regulations. Here’s a possible approach:

HackMD

Over coffee and bagels, I had the chance to talk to some very professional and interested people, e.g. working on international governance - in particular concerning the rights of data workers, innovating legal decision workflows, or building infrastructure to monitor and keep in check the proliferation of AI tools everywhere. We talked about Apertus capabilities, exchanged tips on model configuration and prompt writing. This will be a good venue to propose and vet constructively critical ideas in openly developed AI.

Final prompts

During a chat with Apertus itself, a couple of ideas surfaced as well: issues of data protection, collaboration with research, long-term sustainability. Open questions for discussion: avoiding dependency on large tech corporations, ethical considerations in AI adoption, preparing employees for AI changes, sustainability of AI projects in SMEs. I took home some additional 'tips' for future participations: bring a real challenge, note some key discussion topics, explore potential partnerships, get ready to ask questions and be open to new perspectives.

Thanks very much to the AI Centre & swissICT for organizing, hosting us and sharing a delicious lunch: captured above in MyFoodRepo AI. A discussion for another time, which you can find already in bits and pieces (no pun intended) on my social feed.

Feel free to @respond, let me know if you have any questions or ideas for this community - one that I am hoping to support with my full attention in 2026.

DINAcon is the annual Swiss conference for Digital Sustainability ("DIgitale NAchhaltigkeit"), organized by CH Open and partners, that took place again in Bern yesterday. You can find the official report, photography and (soon) video here:

dinacon 2025 – DINAcon

Digital sustainability ensures that digital assets like data, software, and content bring long-term societal and ecological benefits. One of the main discussion topics of the conference, is how to ensure that digitalization not only progresses, but is also designed sustainably and accessibly. In Switzerland, this affects many areas such as digitalization policy, data protection, copyright, as well as technological developments like Open Data and the Blockchain.

We had over 270 guests from IT, communities, administration, politics, and civil society meeting at the Kongresszentrum Kreuz this year, with 36 presentations on a variety of current Open Source technologies, applications, infrastructures, and legal developments. I am very thankful to the speakers, sponsors and all the participants for a very insightful day.

A fundamental topic again in 2026 was security and data protection technologies to strengthen digital sovereignty, and one way we do this each year is with the DINAcon HACKnight - our after-hours mini-hackathon for Digital Sustainability in Practice. I covered this already back in 2017 on this blog.

Videos and slides will be available soon - I'll update this page in a few days. Here are a few highlights of the conference, which you can also see in my live blogging on Mastodon during the entire day, in a single thread.

Morning

Details on each speaker can be found in the official program (German / French).

Having a strong law in the books is one thing, consistent follow-up is another. From Bulgaria to Trinidad & Tobago, Paula Grzegorzewska from the Linux Foundation talks about going the distance with „First, recognition, next - strategy. On open source in European policy“. Paula referred to the EU digitalisation summit in Berlin, happening in parallel to DINAcon:

EU-Summit - Bundesministerium für Digitales und Staatsmodernisierung

Zur Startseite - Bundesministerium für Digitales und Staatsmodernisierung

During the rest of the day I followed (and tech-supported a few) sessions in the Mobility and A.I. tracks.

In an excellent session on Mobility Data & Services, we heard more about the potentials, role and risks of data spaces in the future of public transport. I was glad to hear about notable data accessibility and access challenges from Benno Staub, Lucie Pannecoucke, Pierrick Servais, and Marcel Rieser.

Developing mobility-as-a-service is not easy, as the many failed platforms highlighted, but having MODI as a feature makes sense for next gen apps. This was inspiration to pitch a HACKnight challenge on the topic.

In a canton where an ambulance trip can last 80km in hilly terrain, the challenge of coordination is clear. A „Next Best“ real-time data system for emergency services is presented by Jakob Böttger (City of Bern, above) that dynamically combines traffic data, deployment patterns, resource availability, and medical needs to automatically select the fastest and most suitable vehicle.

Afternoon

After a quick lunch break, we "Unveil the environmental impact of the ICT sector in Switzerland" – with a talk by Constant Riolo, who draws from a landmark 2024 study by Resilio - in partnership with the ITU, EPFL, Swisscom, SwissICT and many others. Recommendations include enhancement of recycling, BYOD policies, and customer awareness.

The convenience trap of Big Tech is eloquently discussed by Nina Müller (above) on behalf of Nextcloud. An Open by Default policy seems a no-brainer after looking at the thought points presented.

With “the Functional Framework”, like the open API registry from TM forum, companies can put consistent standards, requirements and access pathways for customers or suppliers in one place. Florian Kammermann (above) showed us in detail how this is being used by Swisscom to standardize interfaces among telcos and Service Providers. We also promoted this topic in a HACKnight challenge.

stepping stone provides secure, Swiss-based AI infrastructure and HPC, supporting industrial science projects with complete data sovereignty (US-hardware excepting, as an audience question pointed out) and security requirements. An excellent input from Michael Eichenberger, CEO, which dove into some of the scientific/industrial applications being developed on their open source platform. Really recommend checking out the full talk when it's available.

Inference Utility

On a topic closely related to Michael's input above, in the afternoon I replaced Patrick Hofer (who was unfortunately was out sick), to repeat a talk I gave the day before in French in Biel/Bienne. My input into the AI track covered some history, and shared an outlook for the civil sector. In particular, I discussed the Public AI movement as a way to get behind sovereign ML/AI projects, and build on the foundations of Public Money, Public Code. Specifically I discussed the work of publicai.co on the vision of a public Inference Utility, making the Apertus model available early on to non-technical users in Switzerland and other countries.

My slides - with some ideas for how to get involved - are here:

Inference Utility

Develop large language technologies in the open - as a public service tool, emphasizing sustainability, neutrality, transparency, and individual agency in using AI responsibly.

Dribdatloleg

An earlier blog post on the topic of Apertus is also relevant:

110 #apertus #instruere

Learn about my initial experiences working with the Apertus Large Language Model during the Swiss {ai} Weeks, with advice on getting started yourself.

define:aletsdatOleg Lavrovsky

In the final track session of the day, we heard from Matthias Günther, who gave the low-down on EMBAG Art. 9 - a law that requires the 🇨🇭Swiss federal government to publish commissioned software. This alone generates no direct benefits, he said, going through several historic and current example projects, to draw from deep experience on achieving real synergies and cost savings.

Open development must involve stakeholders from the requirements gathering phase and include collaborative development with selected third parties, requiring adapted governance structures. This was reiterated in a powerful final keynote from a legendary figure in the open source world.

What Freedoms are we losing? What do we really mean by sovereignty? What happens when we click „I agree“ and give up control of our systems? 🫣 Henry Poole showed us the blind spots, with maturity models for service firms, small businesses even, that find a sustained purpose in digital stewardship and technical ecosystems.

Join the FSF 40 Hackathon to celebrate and support four decades of free software activism this weekend! Henry Poole was joined by a brilliant panel for a closing discussion.

Rahel Estermann (speaking above) has put in almost 40 proposals at the Lucerne parliament, several on topics of open data and software freedom. It was a great wrap to the conference with a visionary panel discussion moderated by Clarisse Schröder.

But it was not the finale - we were treated to art and music, delivered by Sophie Dietiker, a secondary school student and talented singer performing „The open world is our legacy!" – an upbeat anthem set to a GenAI soundtrack and generative video that is representative of the unbelievable times we live in.

At last we are at the end of the staged event, and it's time for the Apéro + HACKnight! But first, I need to tell you about our poster session.

Prototype Fund

Sharing the largest room of our location with stands of our sponsors, there was a showcase set up by Opendata.ch to highlight some up and coming initiatives. This lively exhibition of new ideas and innovative technologies took place in the old chapel.

Meeting such lovely people, you might quickly find yourself discussing the challenge of running AI applications on embedded hardware. Also known as 'edge' devices, Tensor Processing Unit (TPU) boards like these pack a lot of punch.

As demonstrated in the real-time computer vision applications of the WildCamera project:

You can find this project and other Round 4 teams of the Prototype Fund here:

Projekte - Runde 4 - Prototype Fund CH

…

Prototype Fund CH

A real highlight for me was getting to know the exhibitors – such as Urs & Miranda from the Zero-Power-Indicator (ZPI) - who also developed this incredibly useful device:

They have a whole bunch of other interesting hardware hacks in their shop:

GaudiShop – Open Technology

GaudiLabs LLC

And we hacked on their documentation during the HACKnight, producing this wiki:

GaudiLabs/qPocketPCR | DeepWiki

This repository contains complete hardware design files for the qPocketPCR portable PCR (Polymerase Chain Reaction) device. The repository focuses on two key physical assemblies—the LaserLid component

DeepWiki

HACKnight

During our traditional "apéro with keyboards", we welcome people and machines of all skill levels, inspired by open hackdays and hackergartens. On this page we post challenges, issues and activities from projects presented at DINAcon. We also provide a quick introduction to anyone interested in learning more about the activities of CH Open. All relevant proposals to start, improve or support open source projects are welcome. The entry is free, and so is the shell 🖪 and beer 🍺 thanks to all the partners and supporters of DINAcon!

During the day, we scribbled ideas and put stickers on a board next to the Prototype Fund exhibition: basically our mini-unconference. The ideas were transferred to a Dribdat site at hacknight.dinacon.ch - some people joined us just for this part of the day's proceedings.

We gathered round in sharing and debating the latest trending topics, such as GNU Guix.

Check out the documentation and logs of these projects if you're curious to see more of what we were hacking on. Don't miss the microposts in the HACKnight Dribs as well.

OSS Catalog

POC based on publiccode.yml

Dribdat

OpenParlData.ch

We have collaboratively built an open API for harmonized data from national, cantonal and municipal Swiss parliaments. It’s currently in beta. Use it to create your own tool, visualization or analysis and promote transparency, participation and innovation in Swiss politics. We welcome your feedback via GitLab Issues (or by email)!

Dribdat

qlever

Graph database implementing the RDF and SPARQL standards. Very fast and scales to hundreds of billions of triples on a single commodity machine.

Dribdat

WildCamera

Open Source Wildlife Camera

Dribdat

Earn Open Data

An open data guide to earning and funding opportunities

Dribdatloleg

Zero-Power-Indicator (ZPI)

Electronic hardware component that only uses energy on a state change; Zero Power consumed when static.

Dribdat

Opening doors to MODI

MODI (Mobilitätsdateninfrastruktur) consists of two parts: “Das Verkehrsnetz CH bildet das Schweizer Verkehrssystem einheitlich ab; NADIM ermöglicht die Bereitstellung und den Austausch von Daten zu Mobilitätsangeboten und Infrastruktur.” Help identify data coverage gaps or opportunities for new mobility services in specific Swiss regions.

Dribdat

Vibe-Helm

A brief investigation into vibe-coding deployment scenarios (e.g. Helm chart)

Dribdat

Logo Generator

the generator creates a logo from your sketch for your buzz/proj.

Dribdat

ODA Canvas

The Open Digital Architecture (ODA) Canvas is an execution environment for ODA Components and the release automation part of a CI/CD pipeline. It provides access to a range of common services (for identity management, authentication, observability etc) and has a set of Software Operators that automatically configure these services based on requirements defined in the ODA Component YAML specification.

Dribdat

Inference Utility

Develop large language technologies in the open - as a public service tool, emphasizing sustainability, neutrality, transparency, and individual agency in using AI responsibly.

Dribdatloleg

Thanks everyone supporting DINAcon this year - stay tuned for full coverage - and save the date!

November 18, 2026

Presentation from my keynote at Hackathon on Crisis Applications, an event organized by the Digital Society Initiative at the University of Zürich on October 30, 2025.

Over three days, students will be working on the creation of crisis response applications addressing real-world problems. As part of the Government as a Platform research project, the project team is organizing a hackathon where students can work on practical problems in close collaboration with the experts in the field. The goal is to come up with new innovative solutions for crisis response to enhance digital resilience of our society. – dsi.uzh.ch

Hackathons as crisis response - GaaS talk - HackMD

View the slide with “Slide Mode”.

HackMD

Since childhood, I've been a fan of programming contests, science fairs and hackathons. This picture of me was taken by a friend in Lausanne some ten years ago. Very much in my element, I'm unpacking open hardware kits to help small teams of people work in a public space on agricultural challenges from a department of the Swiss Federal Government, articulated with Sustainable Development Goals.

Whether hackathons take part over the course of a few hours, or several days, the idea is generally the same: challenges are the input, around which teams of volunteers form to work within strict time-limits on a solution, usually documented and presented in the form of a pitch as the output.

An event where programmers() meet for 
{collaborative computer programming}.

Today you don't need to write code to create software, but in popular awareness, hackathons are still rooted in the rites and rituals of computer programmers. This is also how I also got to know the format, by taking part in the very technical hackathons of Silicon Valley companies and open source communities. We should aim for a more inclusive definition of 'hackathon', that explains civil society engagements from diverse stakeholders.

The relationship between hackathons and crisis response, particularly visible through international solidarity and disaster relief, is an increasingly important topic. Volunteers from around the world work together to map roads, collect data to assess damage, or to support infrastructure critical for aid delivery. Random Hacks of Kindness were set up to coordinate global hackathons focused on disaster response, or even just to support charities with tools for information management and response. These events often prioritize projects that can be rapidly deployed or scaled during crises. They were my first hackathons in Switzerland, and I contributed to the development of tools to locate lost relatives.

Following the Haiti earthquakes in 2010 and 2021, Typhoon Haiyan in 2014, and many other disaster areas, the Humanitarian OpenStreetMap Team organizes global initiatives to rapidly update maps for relief efforts. This week, Hurricane Melissa in Jamaica (data on affected areas pictured above) is in the news as the "storm of the century", leading to outpourings of support aid and interventions. It's quite likely there will be some hackathons to deal with the aftermath, a format that brings institutions and communities together.

Today with the proliferation of corporate advocacy and technology-invested governments, giant companies are often behind hacker relief programs, such as IBM's Call for Code Hackathon to find solutions for natural disasters, pictured in a news article above. Perhaps the motivation here is Corporate Social Responsibility, almost certainly such events (with grand prizes in the tens of thousands of dollars) are not an act of simple charity. Here the crisis hackathon is a highly strategic element of marketing and tech evangelism.

One research hypothesis, is that participants from different cultures and backgrounds contribute to solutions that can be uniquely adapted to certain contexts. In the best case, this helps to foster a culture of global solidarity and mutual support in crisis response. Working with a pilot group of refugee and migrant participants, and their support organizations, the dynamics of hackathon inclusivity are being studied by my colleagues at BFH and UZH on the HackIntegration project.

Seen in events like Versus Virus during the COVID-19 pandemic, or hackathons organized for beneficial causes like Ageing Well and Social Work, a dynamic develops that has been the subject of analysis. Such studies have already shown that open mapping initiatives like OSM can increase the efficiency of disaster response. For example, the World Bank found that OSM data helped in routing aid convoys and identifying areas of need more effectively, though challenges like data quality and validation were noted. I am eager to support research into the effects of hackathons like the one shown above, and am developing my own data aggregation tools while supporting global advocacy.

Most recently, the Swiss {ai} Weeks formed as a broad grassroots alliance around the debate (some people consider them a brewing crisis) on the sovereignty (trustworthiness) and sustainability (ecological and otherwise) of AI technologies. Supporting the Apertus model built to benefit the public good, a dozen hackathons and over two hundred meetups were run on the framework of the Sustainable Development Goals. I cover this in some detail in my blog (utou.ch)

Last week, we had a workshop last week with a small scholarly community addressing such topics of interest at an international level. We shared data across events and borders, involved ethical and peer review, identified gaps and collected ideas for research direction. Visit the Impact section of our shared library for more literature and evidence-based publications. We would very much like to hear from you about any critical thoughts or directions.

In conclusion, I'd like to send you off with three calls-to-action:

💻💡 raw;data;now!

Hackathons and open mapping efforts have been instrumental in mobilizing international solidarity and addressing crisis response challenges. They enable rapid, crowdsourced data collection and analysis, support early warning systems, facilitate collaboration across borders, and question the often inefficient and unequitable status quo.

How will you use the data available to you today?
What is still missing to complete the picture?
What data will you create in the process?

🇯🇲 Support Jamaica

Examples from COVID-19, geopolitical or natural catastrophe highlight the potential of tech-community-driven initiatives to support relief and preparedness efforts. Addressing opportunistic or hypocritical aspects concerns requires a careful ethical engagement, community involvement, and sustainable, context-sensitive solutions. Here are some suggestions, collected with the help of Apertus:

• Openly discuss intentions, limitations, and potential challenges upfront.
• Ensure local stakeholders are involved from the outset and have decision-making power (co-design).
• Plan for post-event support and maintenance, rather than just the hackathon itself.
• Regularly assess impact and be transparent about outcomes, successes, and failures.
• Respect Local Knowledge: prioritize local expertise and solutions over imposed techno-fixes.
• Encourage media to cover the complexities of crisis response and avoid simplistic "tech saves the day" narratives.

Many thanks to all the DSI event organizers and supporters for your interest and our discussions today. Good luck at the hackathon!

Further reading

Choices We Make in Times of Crisis. Waeber et al, Sustainability 2021, 13(6), 3578

Hack for impact – sociomateriality and the emergent structuration of social hackathons. Faludi 2023

Engagement in Practice: Social Performance and Harm in Civic Hackathons. Angela Chan 2021

The potential and challenges of open data for crisis information management and aid efficiency: A preliminary assessment. Stauffacher et al 2011

Europe’s population crisis: see how your country compares – visualised

How anti-immigration politics across the EU clashes with demographic reality

The GuardianAlex Clark

Chaos testing: How crisis drives organisational change

Chaos testing exposes an organisation’s weaknesses, highlighting how crises can drive resilience and innovation.

This is a report on the Swiss {ai} Weeks hackathon in Liebefeld area of Bern, interspersed with visuals, providing detailed results and follow-up references to complete sources, galleries and hashtags. Ending on a brief philosophical inquiry into hackathons as an element of Swiss democracy.

Skip to section:
0. Location
1. Platform
2. Teamwork
3. Outcomes
4. Gratitude
5. Apropos

Switzerland launches transparent ChatGPT alternative

Swiss Apertus LLM aims to compete in a crowded field with openness and accessibility.

www.swissinfo.chMatthew Allen

As mentioned in prior blog posts, the Swiss {ai} Weeks are a pioneering effort – similar to past Digitaltage (Digital Days) and other national cyber-expos. This broad alliance formed around the topic of A.I. built for the public good, and on the framework of the Sustainable Development Goals. Nationally, there are over two hundred happenings in enthusiastic response. As Regional Coordinator in Bern, my focus was to organize a trade fair and hackathon last week. In this post, I dive into the results of these events.

Enter the F{ai}r Hackathon. Free from the lonesome certitude of a user handbook, or carefully planned out requirements document, we make fresh plans, we bootstrap, copy-paste, vibe code, Getting Things Done in a spontaneous, improvisational manner that seems to me more like a jazz band's improv jam session than a button-down crunch session of an engineering team (* we hope that both can learn from one another)

Location

To launch the new hashtag, we printed #siliconlovefield t-shirts and ran the F{ai}R – space was made for a dozen stands, a day reserved for AI presentations from local companies. There were all-day workshops from embedding chatbots into a business model, to coding live agents on a real-time battelship board. We learned how to prompt creatively and effectively in a competitive prompt battle, did a lot of networking, talked shop, and showed off the infrastructure.

Sandro (above) and Selma were amazing hosts for us last week, organizing hot and cold food and drinks at their bar, complemented by Vi’s cuisine in the food truck outside. If you have a chance to come visit "Silicon Lovefield", our new hub in Liebefeld, I encourage you to stop at WORKSPACE & MORE - the most light-filled, greenest, quietest and geekiest coworking space in the region.

The F{ai}R started early on Monday, and we reserved Thursday and Friday for the hackathon - two hackdays connected to the presentations and exhibitions. Some people put in ideas, others came back to also take part in the hackathon, energized with their learnings and questions. You can find more impressions in my previous blog post, or watch the SRF coverage of the event to get a 5 minute impression of the venue.

109 #swissaiweeks #siliconlovefield

A whirlwind of activities generated (no pun intended) by the Swiss {ai} Weeks in Bern.

define:aletsdatOleg Lavrovsky

KI: Programmierer testen das «Schweizer ChatGPT» Apertus - 10 vor 10 - Play SRF

Was kann das digitale Sprachmodell Apertus? Bei einem Hackathon in Bern testen verschiedene Interessierte Anwendungen basierend auf dem «Large Language Model» der ETH und EPFL.

Play SRF

Platform

At the start of the Bern Hackathon, we presented 10 challenges across a range of ideas and sectors. Anyone could submit an idea through our website, or at a series of in-person and virtual workshop sessions in the summer. We received over 20 submissions this way, and also encouraged continuing challenges from previous hackdays. Eight of the challenges were finally worked on by some of the 92 people who registered and created a user profile.

The Dribdat site is where you will find a description of every challenge and resulting project. This is an open source, self-hosted web application developed in support of open data hackathons. Each of the Swiss {ai} Weeks hackathon organizers built a slightly different platform, using Discord, Notion, Craft, and other products. Dribdat itself relied on authentication and storage from international cloud providers. So, as far as I know, no hackathon site used a fully Swiss-hosted solution: something that should be discussed in retrospective.

Our Dribdat instance was hosted by Ungleich in a 100% carbon-neutral data center, connected to Hugging Face for authentication, Linode for object storage, Cloudflare for reliability, reCAPTCHA for spam protection, and Fathom for analytics. You can find the latest source code on Codeberg, which is based on Python, Flask & Bootstrap.

We used the Apertus 70B instruct model provided by PublicAI for generating challenges, tips and evaluations. In the introductory session, we encouraged people to use their choice of collaboration tool, and announced a total of four alternative options for accessing Apertus: including local, national and international providers of AI services. All these, and 25 other technical platforms that were supported at the hackathon, are listed in the Resource wiki.

Habemus Apertus

Complementing the launch events from the Swiss AI Initiative, we had a challenge that was completely Apertus-generated, based on the description of our hackathon on Dribdat. It was spoken out by an AI voice from ElevenLabs, on stage represented by a small Otto DIY open-source robot. That was an enjoyable nerd-moment at the end of the presentations, even if a team didn't form around the idea. Perhaps next year, when we get a more expressive robot as your coach and mechanical team-mate 🦾🤖

It was interesting to see this auto-generated proposal to "consider as a team how to truly meet the needs of global communities by enhancing specific cultural capabilities - accessing untapped datasets, or even advocating for data contributions like it was done already for Rumantsch dialects". It reflected the vision manifest (Appendix O) of the Apertus Technical Report, as well as the system prompts of the PublicAI deployment.

Habemus Apertus

This challenge to tap into public data sources is completely AI-generated.

Dribdatluisdbarros

While several of the mentioned data sources were hallucinated (m4cite ? Erasmus+ ?), the contact information at the end reads like an Easter egg from the designers, inviting hackers to connect with improvement suggestions – the hackathon providing an excellent venue to draw attention and new recruits.

Indeed, the Habemus Apertus challenge was something that you could say was in the background of all our projects. You can read all about the new Swiss-made LLM model in my earlier blog post, including instructions of how to get it running on different services and your hardware. We were fortunate to have the expert guidance at the hackathon of Prof. Marcel Gygli (presentation, slides), who joined us for a tech session on Thursday.

There was no requirement to use Apertus at the hackathon, but we can assume that everyone tried the free chat. Several of the teams explored model capabilities and used it seriously, and there were several instances of proposed improvements. Questions about the reliability of results, provenance of the datasets, hosting and training options for the Apertus model came up. We had a physical copy of the Tech Report on site, a Hugging Face forum, GitHub repo and Discord server, where we could forward tricky questions to the experts.

Prompt Buskers

This was a 'meta' challenge where we aimed to collect and express creative ideas during the event. Note that it was generated by Mistral 24B and inspired by Buskers Bern, a few weeks before we had access to Apertus. We invited CyberGwen, a local musician, who first played at the F{ai}R, then performed for nearly two hours on Thursday evening.

By thinking and planning ahead of time to involve people from diverse backgrounds, bringing in a cultural performance at a critical juncture, it was a small but meaningful contribution to the gathering. I hope that it will inspire you to organize creative workshops, and support local entertainment artists at future hackathons. Watch CyberGwen tune up the night with open source visualizations – full credits are in the video description:

Prompt Buskers

Motivated by the ideas in this challenge, we invited a local musician to perform a set during the Hackathon. This was the first live act of CyberGwen, who has already a growing audience online. The visuals in the background from open source shader artists complemented the act. It was epic. A recording will be available soon.

Dribdatloleg

Teamwork

Our participants were asked to split into groups of up to five, though it was more of a rule of thumb rather than strictly enforced. You can see the directory of 92 users who created a profile on Dribdat. During the check-in (photo above), participants got a recommendation of a team based on the preferences in their profile. Thank you to the BFH / University of Zürich for helping with the team building process! For more information, please visit the interim report of HackIntegration.

Tables were placed in clusters all around WORKSPACE & MORE, two teams worked on tables outside, thanks to the sunny weather. We had paper and digital whiteboards, moderation kits for brainstorming, and enjoyed a speedy Internet: only at one peak moment was the 25 Gbit/s connection from Init7 saturated. Snacks, water, tea, coffee, maté were available from 8 AM until 10 PM. The location was fully accessible, and we had a couple of participants with restricted mobility who appreciated this.

The following 9 project teams presented their results, in this order, on Friday:

1. Energy Infrastructure from Remote Sensing Team Beta
2. Guardrails as Code
3. AI Mates powered by Apertus
4. Archive Image Matching 🏆
5. Create your own Planetary Systems
6. Local produce transportation 🏆
7. Measure footprint of open LLMs
8. Tibetan Chatbot 🏆
9. Energy Infrastructure from Remote Sensing Team Alpha

In the following sections, I review each project in a bit more detail. You can also skip the first 13 minutes to watch the presentations in our full video recording of the presentations, or watch my 20 minute short video review of the documentation.

Energy Infrastructure from Remote Sensing

From this year, residents of the Canton of Bern are not obligated to register their solar panels, and two teams (Alpha and Beta) worked on the problem of getting better insights into energy production. The goal of the challenge, proposed by the environment and energy commission, is to use open data to estimate production capacity, and create tools to help with energy planning and communications.

We had a chance here to learn a lot about the topic of collecting urban indicators, and exploring them with an A.I. system. They first wanted us to work with sophisticated, highly granular datasets, but ended up simplifying things a little bit. There was actually a whole hackathon run one week before ours on the topic of energy data, also part of this year's Swiss {ai} Weeks: the Energy Data Hackdays. I was there and coached one of the teams.

One gets the sense that this is a huge topic: how do we address the climate change crisis with more resilient infrastructure in terms of energy generation and consumption? You can see a little bit in the log that this team struggled to have the right computing capacity. Both teams worked really hard on on the challenge, training a model with a new AI architecture to improve the rate of detection of solar panels, then plotting the results on a map to see how effectively the cantonal policies are being implemented in various regions.

The Alpha team collected and labelled images of roofs in Bern, aiming to train a data model that consistently recognized patterns in energy production facilities. In the Beta team's presentation you can find a Readme (in Italian), which describes how they acquired data from sources like Swisstopo, extracted it and analyzed the orthophotos. Both teams used computer vision techniques - similar to the Archive Image Matching team - to detect the energy infrastructure from satellite pictures. Very interesting research attempts were made, along with supporting visualizations and prototype dashboards.

Team Alpha

Recording from 1:14:00

Energy Infrastructure from Remote Sensing Team Alpha

Estimate energy production from open data, and help to inform cantonal energy planning.

DribdatMonika

GitHub - lucasahli/swiss-ai-weeks: Hackathon

Hackathon. Contribute to lucasahli/swiss-ai-weeks development by creating an account on GitHub.

GitHublucasahli

Team Beta

Recording from 14:00

Energy Infrastructure from Remote Sensing Team Beta

Estimate energy production from open data, and help to inform cantonal energy planning.

DribdatPeterZweifel

GitHub - longobucco/bern-solar-panel-detection

Contribute to longobucco/bern-solar-panel-detection development by creating an account on GitHub.

GitHublongobucco

Guardrails as Code

Technical guardrails are safety measures designed to ensure that artificial intelligence systems operate ethically, safely, and within defined boundaries: functions and practices to help certify that a service is compliant. For example, to help ensure your chatbot or automation stays within more or less strict delineations of what it is supposed to know, chat, or decide about. Interprimis, a local consultancy, challenged us to apply them to AI.

In the case of Swiss and European laws (think A.I. Act) this may be data protection issues, for Internet content you may define a scale of toxicity ratings. The project addresses to various issues they see in implementing guardrails. The team conducted some research, generated a proposal with OpenAI which discusses the impact and costs of implementing - or, indeed, the risks of deactivating - guardrails in a design blueprint. I would have liked to see a small demo, but they lacked a developer in their team. They shared resources, clarified requirements, and expressed a readiness to involve people in the future in addressing a critical topic that is a part of most LLM systems right now.

Recording from 21:30

Guardrails as Code

Urs - Interprimis

Dribdat

AI Mates powered by Apertus

A dating platform idea with a twist! Not just another Tinder clone: here you have an authentic challenge, an audacious idea, a true desire to help people and apply personal experience. The result was arguably the most dramatic and memorable presentation on Friday, with the team acting out their arguments to revolutionize modern romance. The Agentic approach of the solution is the hottest thing in the business right now: in this app, you will create an avatar, sharing your data with a virtual agent that goes out and "meets" others like it, sharing learning about the experiences before you go out on a real date.

The team worked intently on a prototype using the PublicAI and Brandbot instances of Apertus, and came up with a convincing click-prototype. They thought a lot about what it would take to implement it, worked on extensive product requirements. My time as organizer with any of the individual teams was very limited, but in the night during the hackathon I vibe coded a mock MCP (model context protocol) server. We will see soon if this is one of the approaches we can take. On the whole, a nice initiative, clearly something that the whole team is passionate about, and sees a market ready for. Let's AI Mate!

Recording from 28:40

AI Mates powered by Apertus

Authentic Dating made in Switzerland

Dribdat

GitHub - datalets/luvatar-mcp at feat/mcp-matchmaking-server

Contribute to datalets/luvatar-mcp development by creating an account on GitHub.

GitHubdatalets

Archive Image Matching

A project from the University of Bern, that was done by a team of people who wanted help to categorize images in a very large archive of hundreds of thousands of pages of historical prints. In the past, this would have been a popular crowdsourcing task, today the goal is to distribute the work among AI tools. There are many issues and constraints to work with to get these texts read in properly with advanced computer vision, and clustered with LLMs.

The solution was to augment traditional techniques with new algorithms to optimize term frequency and improve the accuracy of the results that they're getting out. They trained a very high performance AI model that takes half a second to process an image. To correctly manage the matched images, they also created a frontend and backend catalog. This is a FastAPI based server connected to Apertus, with a bit of code that allows searching, uploading, and better understanding the content - and lots of cool little features built in. I really appreciated seeing a working demo with Brandbot running Apertus. Very cool that they came up with this, and a big value provided to the university team.

Recording from 37:00

Archive Image Matching

Visual Matching in Historical Print Catalogues

Dribdat

GitHub - xaviermolinaa/Image-Matchmaking-Api--E-rara-

Contribute to xaviermolinaa/Image-Matchmaking-Api--E-rara- development by creating an account on GitHub.

GitHubxaviermolinaa

🏆 Congratulations to the team for winning the Public Vote!

Create your own Planetary Systems

This team worked on an AI algorithm developing an understanding of the structure of exoplanetary systems - planets like our own in other solar systems. In a fun set of presentation slides, starting with this ancestral person looking up at the stars wondering is there life out there? You’ve got to love space science. Lots of data, different telescope arrays and "are we alone?", the deep and searching questions that robots should help us with. We were a rapt audience!

The project was discussed again after our hackathon in the on{ai}r webinar. In the video linked above, they talk about the general topic of exoplanet research, how crowdsourcing helps the discovery of other Earth-like planets, the role of AI, and review the project in depth. Multiple prototypes came out of this this team, both an intense statistical model with deep number crunching and data analysis, as well as an interactive solar system you can play with. Don't miss their Dribdat log to see all kinds of outputs from their hackathon experience that did not make it into the final presentation.

Recording from 43:30

Create your own Planetary Systems

Develop an AI algorithm capable of understanding the structure of (exo)planetary systems and generate others.

Dribdatscm99

GitHub - tjahn/swissai_planetary_systems

Contribute to tjahn/swissai_planetary_systems development by creating an account on GitHub.

GitHubtjahn

Local produce transportation

"Why don't you order your food from a local farm shop?" Starting with this simple question, the team worked to create tools for shops and farmers to easily transport their goods to customers. In their presentation they presented Farmly - even coming up with a logo and brand – to mock up an app. It allows chatting with an intelligent AI service, that accepts tasks and cues them into an ordering platform, where you can order various fruits and vegetables or farm products. The route that these products would take is calculated using the OpenRouteService API.

They built and demoed their solution - mainly designed for farmers, thinking out in lots of detail how it would work. Great results from a team of young people who worked without their challenge owner, as that person unfortunately got sick just before the hackathon. Very courageous of them: and doubly so to make a national television appearance with an in-depth interview in the middle of the hackathon. Hope to see a launch page for the product up and running soon!

Recording from 50:00

Local produce transportation

Create tools for local farmers to easily transport their goods to customers

Dribdat

GitHub - Kirchenfeldrobotics/Swiss-ai-weeks-Hackathon-Liebefeld

Contribute to Kirchenfeldrobotics/Swiss-ai-weeks-Hackathon-Liebefeld development by creating an account on GitHub.

GitHubKirchenfeldrobotics

🏆 Congratulations to the team for tying for top in the A.I. ranking!

Measure footprint of open LLMs

This team set itself as a goal to really understand how Apertus uses energy, and got under the hood to benchmark the LLM, compare it with other models, and recommend strategies for efficient prompting. The energy profile of AI as being one of the top concerns people have in using the very energy-hungry services that incorporate it, it was very interesting to hear what the team had to say about energy consumption – and to what extent we as everyday users can influence it.

We had various infrastructure made available: a large Mac Studio running in a data center was sponsored by Begasoft. There were also two graphics workstations on site, which we had trouble getting to work. The model is still quite new and doesn't easily run on all hardware architectures: see my separate post for more detail.

110 #apertus #instruere

Learn about my initial experiences working with the Apertus Large Language Model during the Swiss {ai} Weeks, with advice on getting started yourself.

define:aletsdatOleg Lavrovsky

The experiments that the team ran were to test different prompt and response lengths, different subjects, a number of different languages, etc. They ran comparisons with Llama, another popular open source model, to try to really understand how much energy is being used in different use contexts. Unfortunately, they were not able to get a very stable setup. For a good benchmarking, you have to be able to reproduce the results over and over again. They were very forthcoming with their failures: the basis of future efforts.

The team figured out how to measure a range of different consumption values through local and remote access. They clearly learned lots, had fun, created graphs, and shared everything in their project page – which is all great to see at a hackathon. Some very good inputs overall from a team that took a difficult subject, tackled it, and ended up with something that we can learn from and use going forward.

Recording from 58:20

Measure footprint of open LLMs

Benchmark Apertus, compare with other models, and find strategies for efficient prompting.

Dribdatstefan

GitHub - luisantoniio1998/Measure-footprint-of-open-LLMs: Measure footprint of open LLMs

Measure footprint of open LLMs. Contribute to luisantoniio1998/Measure-footprint-of-open-LLMs development by creating an account on GitHub.

GitHubluisantoniio1998

Tibetan Chatbot 

Again focusing on the use of the new Swiss LLM, which states as one of it's main features a strong linguistic ability, the challenge here was to build a RAG with new multilingual references. This team really applied itself to feed in new data from the Tibetan language, working with Unicode text that they got out of a large spreadsheet (open data available on GitHub), or other resources found online. They prepared this to create a chatbot for language learners.

The team used the Apertus 8B local model, finding that it works no less well than the 70B for the specific requirements of their app. They tested various configurations including local deployment, showing the edge computing potential of more lightweight (and less energy-intensive!) AI models. Their demo using a Streamlit frontend was demonstrated live in the hackathon presentation. A screencast is available on their project page.

The idea of having an AI tutor is one we all understood, and it was to see a live demo that we could play with during the hackathon. The Tibetan learning community will hopefully get to benefit from these additional tools. The fact that Apertus is in the center of attention here is a great call to action, seeing whether the model delivers what it promised. I thought this a promising project, that we are going to be hearing more about soon.

Recording from 1:06:00

Tibetan Chatbot

To help learn a new language using chat interfaces, let’s use Apertus to build a RAG based on a dataset of linguistic references.

DribdatTenzinJhope

GitHub - TenzinJhopee/Lhayum_alpha: Lhayum chatbot - Tibetan chatbot

Lhayum chatbot - Tibetan chatbot. Contribute to TenzinJhopee/Lhayum_alpha development by creating an account on GitHub.

GitHubTenzinJhopee

🏆 Congratulations to the team for tying for top in the A.I. ranking!

Outcomes

It's hard not to be impressed with the energy and dedication of a professional hackathon. From the moment the doors opened until the final project pitches closed, members of every team took the opportunity to get fully engaged, sharing ideas and working together with remarkable camaraderie. Nine excellent presentations showed evidence of a truly cooperative setting, encouraging everyone to pitch in.

What impresses me the most was the community spirit—whether it was teams solving complex challenges side by side, learning new skills from each other, or supporting one another's ideas. A cooperative energy buzzing on open platforms that keeps us organizers going, and inspires future events. The excellent location and catering is a key ingredient in this, so please get in touch if you'd like to run your own hackathon in Liebefeld!

Some key statistics:

Evaluation

To add some depth, we assessed all the projects using an AI-driven evaluation based on the five criteria of other Swiss {ai} Weeks hackathons: Technical Functionality, User Experience, Skillful use of AI, Uniqueness / Creativity / Fun Factor, and Potential / market impact — all scored autonomously by Apertus 70B. These evaluations were based purely on the project documentation, offering an impartial complement to the public votes. You can find the Apertus reviews in the Log (Dribs) of every project. If you are interested in AI evaluation of ideas and presentations at your event, let us know.

Coverage

In addition to our social media hashtags (#swissaiweeks #siliconlovefield), the event was covered by SRF, Switzerland’s national broadcaster. We have uploaded a photo gallery for a visual record of the hackathon atmosphere. Everyone who participated could get stickers, shirts, and other goodies from the sponsors. We printed a heavyweight paper certificate for every participant who wanted one, available digitally on demand. Several people received recommendation letters for their schools, workplaces, or social services. We presented the hackathon and one project (Tibetan Chatbot) again at AI+X Summit in Zürich.

Awards

These three projects earned special recognition, with team members receiving small prizes (books, noisemakers from teenage engineering, Bern vouchers, tickets to Uphill Conf)

Top in Public voting:

• Archive Image Matching: A project that leveraged AI to help sort and organize historic images, enhancing accessibility and search capabilities within archives.

Top in AI evaluation: (tie)

• Local Produce Transportation: A chatbot that aimed to help local farmers find convenient transportation solutions using AI-powered matching systems.
• Tibetan Chatbot: Developing an AI tutor based on the Apertus model to support access to content and learning of the tibetan language.

To learn more, click through the project descriptions and presentations on our event page to see the full scope of the results. Discover the 27 guides we assembled in the Resources section with explanations of accessing tools from Apertus and Supertext, detailing how they could be used during the hackathon. Feel free to reuse everything under a Creative Commons license for your own event.

Gratitude

Thank you for supporting us - whether you joined us at the F{ai}R & Hackathon in #SiliconLovefield, were part of another event in Bern or another of the 24 cities represented, cheered us on from the social network sidelines, or even stopped by to read this blog post and other coverage! The Swiss {ai} Weeks were a massive undertaking, and it will take time to fully process all the impact.

Special thanks to Selma & Sandro for being our brilliant hosts at WORKSPACE & MORE. We are all grateful to Prof. Dr. Marcel Gygli (BFH) for the subject matter workshop, and to CyberGwen (YouTube) for an incredible concert on Thursday. To the Mê food truck for delicious vietnamese cuisine, and Harry Stitzel (SRF) for intrepid reporting on Friday. To Kim Chai Ly, Jürg Stuker and Pascal Melcer - high five's for critical support at key moments. To all of you who put in your time, energy, human openness and intelligence: your contributions will be remembered.

The event was made possible through volunteers backed by the Economic Development Agency of the Canton of Bern, with additional financial contributions from fers stiftung and Puzzle ITC. Technically, our AI hackathon was extremely well supported by the Swiss AI Initiative, Begasoft, Swisscom, Elevenlabs, Supertext, Hugging Face, Init7, Ungleich / Dribdat and PublicAI.

There are people in the background without whom none of this would have happened - or at least not for a long time: big shout-outs to Sabine Wildemann, Diana Engetschwiler, Daniel Dobos, Christoph Birkholz and team, investing massive efforts to bring people together who want to shape the future together.

Once more: thank you to everyone involved, from organizers, participants, and judges to the university and media partners. Whether you were there or watching from afar, I hope you found the Swiss {ai} Weeks scene to be vibrant, open, innovative, and full of promise. Our week here was packed with exchanges, inspiration, a possibility to test fresh ideas hands-on, to meet people from Bern, all over Switzerland, and around the world.

Links to recordings and reports

Please get in touch if there's anything I've missed!

• Team Presentations
• Pitch Workshop - SLIDES
• Expert Workshop - SLIDES
• Opening & Challenges - SLIDES
• SRF 10vor10 Report (in German)
• Photo gallery - Swiss {ai} Weeks
• Dribdat ⬡⬢ - our hackathon platform
• LinkedIn - #swissaiweeks #siliconlovefield
• Instagram - #swissaiweeks #siliconlovefield
• Mastodon - #swissaiweeks #siliconlovefield

If you are interested in more events like this, use the Hackfinder at hackintegration.ch

Apropos

As this was likely the first public hackathon ever organized in Liebefeld, run during the first edition of the Swiss {ai} Weeks, in support of the historic launch of a Swiss sovereign AI model ... I would like to footnote with a few words about the format, a recurring theme of my blog. To explain why thousands of people join these events every year, one may ask: why do we hack at a hackathon? In reference to the first definition of hacking as "playful solving of technical work", a creative expression that is related to a civil liberty, then we may ask, what kind of freedom do we seek at a hackathon. Is it to be:

• Free from obligations to schools or employers or families? Often we can only participate after negotiating the time-out as an educational or otherwise valuable experience. Certificates and reference letters are valuable confirmations of this.
• Free from having to represent something? People often identify socially, wearing clothing, or having something on their badge, a logo of an employer or a community. We collect stickers (ideally hexagonal), proudly representing our contributions.
• Free from loyalties to any institutions? Typical challenges have institutional provenance, being provided by sponsors or academia. Furthermore, hackathons usually propagate tools and technical frameworks maintained by various institutions.
• Free from expectations of behavior? Overground hackathons can only function when people are respectful of the hosts and organizers, of the law of the land, of ethical principles, Codes of Conduct, and apply general common sense.
• Free from overarching principles? The Swiss {ai} Weeks and many other prominent hackathons are aligned with sustainability goals. Goal 4: Quality education, and Goal 9: Industry, Innovation, Technology and Infrastructure, are frequently cited.
• Free from technical debts? For Ward's sake, don't get me started ...

Let us contrast the experience with another widespread practice. On this last election weekend, a momentous one for digital Switzerland with the narrowly accepted e-ID referendum, I was called in to help count up parliamentary votes in my commune. About a hundred of us were told exactly where to sit and what to do. While on duty, nobody was allowed to leave the room. We were instructed with precision in an atmosphere of speedy purpose. Everything one did got checked and cross-checked, in a sturdy ritual of communal validation.

One may suppose that the absence of such controls gives one a freer sense of agency. You could even think of the civic hackathon as the antithesis of an electoral commission. However, I would argue, that in some ways they are complementary. A public process to evaluate technological boundaries, file issues, voice concerns, vote with your attention and time-commitment, should also be seen as a pillar of an advanced society. Consequently, the hackathon setting needs a safe space as well as rigor and structure to function – more than just the anonymity, the absence of controls, or precise obligations alone.

While hackathons and elections may seem like polar opposites – one, a space of chaotic creativity, the other a model of structured governance – they underline a deeper need about having a framework that allows for agency and accountability. In open innovation, the rules are there to protect the freedom to explore and create, just as in democracy, the rules ensure that every voice is heard and every vote counted. Both celebrate the human desire for self-expression and community, the freedom to participate and support ideas, albeit in different contexts. The clock ticking towards a deadline being the universal constraint with which hackathons, and democratic elections, make their mark.

As a technologist involved in competitive programming for some 30 years, I would say the most important freedom here is that of software freedom, technical choice, the hacker ethic: that I can stick to what I know, or adopt new alternatives. Use well established methods, or walk the path less travelled. At an open community event, I may explore and criticise faults within a supportive peer group, rather than as a lone bounty hunter. In a more closed one, I may need to sign a Non-Disclosure Agreement first. An ability to include a diversity of people while balancing such freedoms and constraints make the hackathon a unique venue – not just the free drinks, food, space and swag.

Fairness Finder

Develop a prototype app that assesses and promotes fairness for data workers (in participative projects, volunteering initiatives, online gigs), leveraging the FINDHR toolkit to ensure non-discriminatory practices.

Dribdat

In this respect, here is another one marked on my calendar: the 40th anniversary hackathon with the Free Software Foundation (FSF), two months from now: "Free software projects and hackers at any stage of their development are invited to participate." .. Nota Bene: "The use of machine learning, like Copilot, ChatGPT and the like, is not allowed." A reminder that not everyone seems freedoms the same way, but that we are free as a society when we exercise our rights, and take part in the debate.

For now, dive into the open projects, share your thoughts. Let’s continue the conversation on our social channels. Pick up an open challenge. Stay curious, stay cautious, stay freedom-loving, stay creative .. let’s keep building the future of AI in Switzerland together: see you at the next hackathon! Send me a reply (oleg @ this domain) if you want my help to run one in your commune.

[Apertus was used sparingly to support the composition of this blog post]

Continues from my previous post, where some history of the development of open source AI projects was followed by discussion of the new Apertus frontier model from Switzerland. Here is a summary of this article by Apertus 8B: (*all spelling in context!)

The article covers the aperture model (built by the Swiss AI Initiative with an open-source ethos), its capabilities, setup options (cloud or local with hardware considerations), and the importance of community input for trust and governance.
It also touches on broader questions about open data, model transparency, and the Swiss cultural/national context of the model's launch. (see Thoughts on Open & Closed)
Practical tips for developers or researchers on downloading, using, and setting up the model are provided along with resources (e.g., in the cloud, Ollama, Hugging Face, and vLLM) and explanation of costs.

From the slides of the Apertus Tech Sessions prepared for the Swiss {ai} Weeks, we are clearly reminded that the goal of the project is to 1) "Develop capabilities, know-how, and talent to build trustworthy, aligned, and transparent AI." and 2) "Make these resources available for the benefit of Swiss society and global actors" - i.e. nowhere does it say that we should expect a production-ready service. The sessions note that our goal here is to help create an open development ecosystem – especially as open models approach closed model in performance over time.

We have seen charts like these, and want to join into the fray:

What we know is that the Apertus model was trained on the Alps supercomputer, operational at CSCS since September 2024, a data center of over 10'000 top-of-the-line NVIDIA Grace-Hopper chips, with a computing power of 270-435 PFLOPS, reportedly in 6th place (June 2024) globally.

Here is how Apertus compares 'on paper' with similar models:

With a basis of approximately 15 trillion tokens, a LOT of data has gone into the preparation. Particularly noteworthy is the high proportion of non-English data (40%) and coverage of over 1,500 languages, including rare ones like Romansh or Zulu. The data was ethically sourced - without illegal scraping, respecting robots.txt and copyright requirements. While this limits access to certain specialized information, CSCS emphasizes: «For general tasks, this doesn't lead to measurable performance losses.»

The Evaluation section of the Apertus Model Card, Section 5 of the Tech Report have various data on evaluations, and I recommend the blog posts at effektiv.ch for a good overview. In the following sections, I will focus on getting the model up and running for your own testing.

Apertus in the cloud

There are already several good options for playing with Apertus with a minimum of set up, as I have outlined in the Resource section of our Bern {ai} Hackathon last week:

PublicAI

The main chat app for the Public AI Inference Utility, based on OpenWebUI

Dribdatloleg

Swiss AI Platform

Enterprise-level support for AI services from Swisscom

Dribdatloleg

Brandbot

AI-Platform - powered by BEGASOFT

DribdatTomvona

Hugging Face

Share your Spaces, Datasets and Models in the world’s largest model zoo, official partner of Swiss {ai} Weeks

Dribdat

With kind thanks to the three providers above, we managed to have a good start last week. Our hackathon platform swissai.dribdat.cc (Dribdat) is connected to the PublicAI API, whom we thank for the free service – generating evaluations for all the teams, like this:

We also added a RunLLM widget (kind thanks for sponsoring a free agent) for user support, and potentially comparison with Apertus:

Would you like to run your own LLM, on premise or in the cloud? Then we need to have a quick talk about the money.

If I were GPU-rich ...

On my Hugging Face profile, you can see the hardware that I brought to our Bern {ai} Hackathon last week: two rather average workstations, at current market prices around 700 CHF each. They are representative of what most enthusiasts could afford. 23.24 TFLOPS is actually not that bad, as far as value for money goes.

The workstation we borrowed for the Data Hackdays in Uri (shown above) was significantly more powerful - costing a factor of 2 - 3X more than the builds above. Nevertheless, it comes at less than half the TFLOPS - reminding us that these values are only a rough approximation of true performance. Read see the project report of our local AI installation in Uri here (in German).

In my tests, as you will need to reserve some memory for your operating system and programs, at least 16GB of VRAM and ideally 20GB should be available to run the smaller Apertus 8B model, as it was provided on launch day. In other words, you would need to try to get a hold of a top-of-the-line graphics card.

This could be the NVIDIA RTX 4000 SFF Ada pictured above, currently retailing at around 1150 CHF, or the RADEON RX 7900 XTX which is about 800 CHF – though note that support for Radeon chips can be a bit patchy. Oh, and even then, don't expect to get much performance out of it: you probably want to get 2 or 3 of such cards in an SLI setup 🤑

Comparing Apples to oranges?

Given the situation above, it is understandable people feel that the most cost-effective way to get the Swiss LLM in house today is Apple hardware. The current Mac Mini with an M4 chip and 24 GB unified memory should be enough to run Apertus, and retails for < 900 CHF at the moment. Going up up up to 12'000 CHF for the behemoth 512 GB version.

Clearly we are not comparing apples to apples here: the way that NVIDIA, AMD and Apple measure their GPU cores and ALU units differs. Your performance may be quite different depending on the way your platform is set up, and models themselves need to be optimized to run decently on Mac hardware in the first place.

And you have to ask yourself: just how hot do these things get? The performance of a powerful Mac Studio (192GB Unified Memory) was evaluated at our hackathon in the Measure footprint of open LLMs project. Here are two charts excerpted from their report:

There are plenty of installation guides online for Llama, that you can also use to install Apertus on your Mac. I have particularly heard good things about the combination of LM Studio with the MLX quantizations. Cool.

There is a lot of debate out there about what constitutes AI-level hardware, and certainly the push to sell new computers is one of the major factors in the global race to build capacity. Just for fun, here is how my Hugging Face profile would look like if I had 4000 Grace-Hopper units, which were used to train the Apertus model 🤗

Whether you are GPU-poor, or GPU-rich, what software do you need to work with Apertus? I will discuss Ollama and vLLM in the following section. Others have reported good performance with LM Studio (Macs) or Lemonade (Ryzen). Another client I have been recommended is Jan.

Open source ChatGPT alternative that runs offline - Jan

Jan is building Open Superintelligence. It’s the open-source ChatGPT alternative that leverages the best of open-source AI.

Open source ChatGPT alternative that runs offline

Using Ollama

Models downloaded from the Ollama library can be configured and managed most easily with an elegant chat interface of Open WebUI. This is my default option, as I have run a shared server in the office and at home for over a year with open weight models like Llama, Olmo and Qwen. The information here also applies to other software based on the llama.cpp library, such as the popular LM Studio.

What is Ollama? Understanding how it works, main features and models

Ollama is a tool for running large language models locally on your system. Check out this article to learn more about its features and use cases.

Hostinger TutorialsAriffud Muhammad

LM Studio - Local AI on your computer

Run local AI models like gpt-oss, Llama, Gemma, Qwen, and DeepSeek privately on your computer.

LM Studio

Due to some bleeding edge methods used to engineer and parametrize the model, we had to wait for a new release of the library and software.

[Updated] As of October 20, the 0.12.6 release of Ollama includes a new release of the library with the needed support. You won't find Apertus in the Ollama library yet, and indeed the standard model does not yet run, as GGUF (a popular binary format for models) support is lacking. However, there are several community remixes of Apertus available that have you covered, such as 8B releases by bartowski, redponike & unsloth, all tested by me with (on first glance) sensible outputs.

Even with an 8GB consumer-grade video card, it is already possible to run Apertus at home. To get started, just run a command like this in your console where you have Ollama installed:

ollama pull hf.co/bartowski/swiss-ai_Apertus-8B-Instruct-2509-GGUF:Q4_K_M

Some people have reported strong hallucinations and high CPU usage, so your mileage may wary. The upcoming version of llama.cpp should improve performance.

Feel free to share your experiences with the community in the discussion area.

Find more quantized versions of Apertus in GGUF, FP8 or MLX formats on Hugging Face. You may still need to write a Modelfile, include the chat template and other bits for a complete model specification if you get errors like this:

I am involved in the discussion and testing nightly releases: stay tuned!

[Model Request] Support new Apertus model · Issue #12149 · ollama/ollama

This is a new model from the Swiss AI initiative. It currently does not load due to Error: unsupported architecture “ApertusForCausalLM” Some tips on getting Transformers updated on the Hugging Fac…

GitHubloleg

See also the CH Open & BFH workshop described in my previous blog, with links to a video and slides where the ecosystem around these tools is discussed:

109 #swissaiweeks #siliconlovefield

A whirlwind of activities generated (no pun intended) by the Swiss {ai} Weeks in Bern.

define:aletsdatOleg Lavrovsky

Using vLLM

This is the route recommended by the Apertus team, and was available from day 1. Based on this, PublicAI and other providers have launched their inference services.

vLLM is a fast and easy-to-use library for LLM inference and serving. Originally developed in the Sky Computing Lab at UC Berkeley, vLLM has evolved into a community-driven project with contributions from both academia and industry.

For a software developer, the library may be easy-to-use, but deployment is not simple. Nevertheless, for an IT team the multiplatform deployment and integration with DevOps tools like Kubernetes would make a lot of sense. Once you have the NVIDIA libraries and CUDA tools set up, a relatively simple way to start vLLM is with Docker.

I have prepared a script, tested with a basic 16GB GPU machine on Linode: the RTX4000 Ada Small, which costs $350 CHF per month ($0.52 per hour) in my region (Frankfurt - Germany). If there's interest from the community, I will put the recipe into a StackScript. Note that you need to agree to NVIDIA's licensing conditions to use their proprietary libraries. My scripts can be downloaded on GitHub, or here:

You should also create a .env file or pass in a couple of environment variables otherwise:

• Set HF_TOKEN to a token you generated on your Hugging Face settings, making sure to allow "Read access to contents of all public gated repos you can access".
• The HF_MODEL parameter should be set to swiss-ai/Apertus-8B-Instruct-2509 - or to any other model, or remix of Apertus.

Note that I'm using the 'nightly' version to make sure the latest Transformers library is used. I've also set max-model-len to a low 4K (the default is 64K, you probably want to use at least 8K), which you can increase if your system allows it.

Using Hugging Face

Most of my initial experiments are in a Hugging Face Space, mirrored in a Codeberg repository, which can be used by developers during our hackathons. It has a typical completions API with the recommended System Prompt, as well as some optimizations and other ways to query the model using the Transformers library.

fastapi-apertus

Apertus (Swiss LLM) hosted on FastAPI with Hugging Face transformers

Codeberg.orgloleg

In addition to the space, I've prepared an Apertus 8B Instruct 2509 endpoint which you can use with a Hugging Face token. Thanks to support from the HF team, this can be easily deployed on your own by entering the model's name at endpoints.huggingface.co:

Make sure to configure the auto-sleep options to your liking. Too short, and you'll be frustrated by the long startup times. Too long, and you may be surprised by a large credit card bill (though you can also set spending limits in the Billing section).

Many thanks and kudos to Leandro and team for their tech support, and giving out free credits, stickers, badges and other goodies to participants of our hackathon. I really encourage everyone to check out their learning resources online:

Hugging Face - Learn

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Learn

What's in a name?

If I may diverge from strictly technical topics for a moment. Using a Latin word in the name is a quaint choice, but not extremely original. There are several companies with it in their trademarks, so some legal discussion will surely be necessary. The relatively inactive apertus.org project is an open hardware camera. This is alluded to by the generative image pasted at the top of this blog you're reading. There is also the Apertium free/open source machine translation platform, but we don't need to talk about that now.. <ahem />

"Professor Nümmerli" made a quite humorous take on the subject (a fondness for Swiss German humor is recommended) in a video posted by comedian Mike Casa on Saturday.

#ai #apertus #apertus #apertus #geneva #funonsaturday #swissai #swissaiweeks #generationgenerative #dontlie2urai | Daniel Dobos | 14 comments

🇨🇭📰🤖 BREAKING: Prof. Nümmerli elected first Swiss #AI Model #Apertus Ambassador? With new voice interface: Apertus? :: ***HOI*** Following his great performance at the ‘inofficial’ launch of the #Apertus Swiss AI Model press conference, the nomination process for Prof. Nümmerli started immediately and is about to conclude soon ... right after: - the collection of internal signatures - the launch of the referendum - then the vote again - followed by the counter-proposal - and the final blah, blah, blah so practically he is already almost fully approved yet, weischt? Pollings show that already 3️⃣0️⃣%, so *more* then the 3️⃣9️⃣% necessary for approval, of #Apertus researchers, developers, trainers and deployers were laughing - so an overwhelming absolute majority. Thanks Imanol, Martin, Antoine, Marcel, Alex, Melanie, Martin, Sarah, Joost, Oliver, Giuseppe, Bettina, Maria-Grazia, Ido, Barna, Angelika, Eduard, Nikodem, David, Adriano, Andrei, Sabine, Katka, Christoph, Jürg, Claudio, Marc, Thilo, Anna, Allen, Dino, Joshua and many others, please consider supporting Prof. Nümmerli’s nomination. Thanks Mike Casa & Mike Casa Comedy for the good laugh - it made my Friday evening after an overwhelming week and looking forward to see you on 19th Sept, in #Geneva. Can we please get the ***HOI*** sound-bit for a voice chatbot interface to be developed during the Swiss {ai} Weeks? Will you join? #FunOnSaturday #SwissAI #SwissAIWeeks #GenerationGenerative #DontLie2UrAI | 14 comments on LinkedIn

LinkedInDaniel Dobos

It's a good name. One can get used to it.

Thoughts on Open and Closed

Upgrading from existing tools and systems to new ones, especially when it involves innovative AI models like Apertus, often requires a careful reconsideration of old and new. I get the feelings of responsibility, a sense of stewardship over projects embodying the values of openness and community empowerment.

The choice of the Apache 2.0 license suggests transparency and intention to foster open & global collaboration. However, we also need active, public discussion around the choice of licenses and governance models. The rhetoric around openness can sometimes mask real concerns, so understanding the logic behind this choice and whether it aligns with community principles is key.

The decision to publish Apertus initially on the Hugging Face site, for which there is currently no direct equivalent in Switzerland, plus the absence of any mentions on SUPSI, CERN, or other universities, has been somewhat conspicuous. It seems to me that the gated (registration required - the "You have been granted access" in my screenshot at the top) initial publication of the Apertus model has provoked the most skepticism.

Logicians Find a Genie

A philosophy webcomic about the inevitable anguish of living a brief life in an absurd world. Also Jokes

Existential Comics

I chatted about this with Apertus, and got some sensible action items to foster trust:

1) Involve the Community in Key Decisions: Engage with open data advocates early and often. Open forums or public consultations on model development, governance, and licensing can address concerns proactively.

2) Transparency in Data and Governance: Publish more documentation on data sources and training processes. Include explanations of the decision-making process around the choice of license and motivation for the current deployment strategy.

3) Strategic Partnerships and Affiliations: Explore and communicate about whether there are less visible partnerships or affiliation that would explain the current institutional landscape.

4) Open Data Interoperability: Consider discussing integrations or complementary strategies with Opendata.swiss, Zenodo, and similar open data platforms to enhance accessibility and visibility but also value governed, controlled access.

5) Highlight Accessibility and Fairness: Ensure that while the main model is accessible through a controlled portal, there exists a pathway for researchers and developers to understand, audit, or build upon the model with clear guidelines and support for responsible use.

Speaking of involving the community: have you signed up to a Swiss {ai} Weeks hackathon yet? (This clip is also on Vimeo)

Who’s behind the 136 partner organizations and why did they join Swiss {ai} Weeks? Let’s find out. Today: Datalets with Oleg Lavrovsky Meet Oleg (latest;-) on Sept 18/19 at the Bern Hackathon at… | Swiss {ai} Weeks

Who’s behind the 136 partner organizations and why did they join Swiss {ai} Weeks? Let’s find out. Today: Datalets with Oleg Lavrovsky Meet Oleg (latest;-) on Sept 18/19 at the Bern Hackathon at WORKSPACE & MORE, co-organized by Canton of Bern. With this post, we also want to say: Thank you, Oleg, for your dedication to the Swiss {ai} Weeks. It’s incredible what you’ve built for the AI ecosystem over the past months!

LinkedInChristoph Birkholz

Connecting to Apertus

It's reasonable to take a step back to ensure that the model is indeed what it promises to be. Legitimate questions about the dataset's origins, compliance with legal and ethical standards should be addressed. Understanding the model's capabilities, the intentions and process behind its creation is fundamental, not only to trust but also to meaningfully engage with and potentially contribute to its roadmap. Here are some places where this is happening:

GitHub - swiss-ai/Apertus-Generation-Issues-Reports: This is the repository for reporting issues with the SwissAI Apertus Model family generation

This is the repository for reporting issues with the SwissAI Apertus Model family generation - swiss-ai/Apertus-Generation-Issues-Reports

GitHubswiss-ai

swiss-ai/Apertus-8B-Instruct-2509 · Discussions

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Apertus

Learn about the new foundation model from the Swiss AI Initiative.

Dribdatloleg

Further helpful references

Apertus: 4 ways to try out Switzerland’s New AI Model · Blog · Liip

Liip is a Swiss digital agency developing web and mobile applications, designing user experiences, and crafting content.

LiipJosefKruckenberg

Apertus is here: What can the Swiss LLM really do? - effektiv

In July we asked: «Can the Swiss LLM keep up?» – now the first generation is here with Apertus. Time for a sober reality check: Where does Swiss AI really stand, and who is it interesting for today?

effektiv

Why choose #Apertus over Llama 3.1 for AI applications | Marcel Salathé posted on the topic | LinkedIn

Still running your AI application on models like Llama 3.1? Try #Apertus - same muscle, no legal hassle. Every week, I see AI demos using models like Llama 3.1. Why are people still building with legally questionable “open” models? “Well - there aren’t really any alternatives”, some of you might say. I ask you to reconsider. The new #Apertus models are both highly performant, legally compliant, and have Apache 2.0 licensing. If you’re building AI applications, this matters more than you might think. If you’re building with such a model, you’ll get: ✅ Simplified legal review and procurement ✅ Freedom to use outputs for training other models ✅ No usage restrictions or special permissions needed ✅ Standard OSI-approved license your legal team already knows 💪 But is it strong enough? Not only is #Apertus highly competitive, it even edges slightly ahead of Llama-3.1! Real-world performance will vary by task, but it’s clearly competitive with leading “open” models. And note that many of those so called “open” models are not actually open in the broad sense. Here’s what you get with #Apertus: 1️⃣ Full training pipeline transparency - scripts, data, intermediate checkpoints all public 2️⃣ Respects opt-out signals - including retroactively 3️⃣ EU AI Act documentation included for compliance workflows 4️⃣ Auditable from start to finish - rare at this scale 🫣 The alternative: Do I really need to list all the legal cases brought against some of the model providers? The dirty (not-so-)secret everybody knows: the leading models have trained on data obtained illegally. It’s one of the reasons their models are so strong (the so-called compliance gap). You may argue that this is not your problem as a user, and I fully understand you. But if you are building, and you are building in a legally sensitive setting, you will definitely need permissive licensing and full auditability for compliance/risk teams. With Apertus, you get that, along with its strong performance. So, if you currently have an application running with something like Llama 3.1, then definitely trial Apertus - developed by EPFL, ETH Zurich & CSCS. Get the models: https://lnkd.in/eZxGHJvN | 19 comments on LinkedIn

LinkedInMarcel Salathé

Apertus tested: How the multilingual AI model performs

With Apertus, Swiss researchers have released an open-source and transparent large language model that cannot catch up with the frontrunners, however.

heise onlineDr. Christian Winkler

Swiss AI’s Apertus 70B and 8B: A Complete Deep Dive into Switzerland’s Revolutionary Open Language…

In the rapidly evolving landscape of artificial intelligence, Switzerland has made a groundbreaking contribution with the release of…

MediumSai Dheeraj Gummadi

As mentioned in my Apertus post, I am helping to produce the Swiss {ai} Weeks – a first edition of a new national platform, as hackathon organizer and regional coordinator for Bern. Here I would like to share some impressions of activities that I was able to join, from kicking off with the on {ai}r webinar, to the first trade F{ai}R in #siliconlovefield

Allow me to begin by extending my deep gratitude to all the event organizers, sponsors, friends and partners – in particular, to the Economic Development section of the Canton of Bern as the first of the key supporters - and the whole #SwissAIweeks community.

Swiss {ai} Weeks

With the theme is "AI Made in Switzerland – Shaped by You", this is a new national platform, that aims to bring more diverse voices to the table of critical debate and knowledge sharing on the opportunities and issues of Artificial Intelligence.

Join the movement! Swiss {ai} Weeks calls on researchers, developers, businesses, and citizens to come together and build the future of AI — hands-on, ethical, and open. This isn't just exploration, it's collaboration in action.

The initiators are the EPFL AI Center, ETH AI Center, Swisscom, Impact Hub Switzerland, Kickstart Innovation, Panter, Wildcard Media, a similar constellation to the VersusVirus events of 2020. I joined as Datalets to help get things running in Bern, while helping out in other locations around the country as well. At the end of spring, this was just an idea, we had no events and just tentative dates. At time of writing, 222 events in 27 cities, 19 events in Bern alone, have been organized. There are 8 hackathons in total joining the parade on swiss-ai-weeks.ch

To get the word out, I designed tri-lingual posters, brochures, pamphlets and stickers using Inkscape at Effinger (with thanks to the coworking community for early feedback). Hundreds of them were printed on recycled paper, hand-cut and distributed around Bern. They were dropped at existing tech/society hotspots: BATbern, Guild42 and DIN meetups, Impact Hub and BFH events, at the Polit-Forum and the Living Room. This complemented the massive digital promotion from Sabine Wildemann & team.

1.9. on {ai}r

How do large language models work and what role does digital sovereignty play? This was the kick-off presentation during the on {ai}r interview series that was run by WldCrd Media throughout September. Here we got a brief introduction to LLMs and saw some of the work being done by researchers in Bern, such as Niklaus et al 2023:

There was a sold-out meetup at Zühlke of the Machine Learning & Artificial Intelligence Bern group that evening, the slides from which can be found here:

meetup/Meetup_20250901 at main · mlbe/meetup

Contribute to mlbe/meetup development by creating an account on GitHub.

GitHubmlbe

2.9. Impact Hub Bern

On the morning of September 2, the Swiss AI Initiative launched the Apertus model - at this point the Swiss {ai} Weeks were really on! You can read about this momentous occasion in my separate blog post. After a morning of downloading and tensor-loading, I supported a workshop on AI for Impact Entrepreneurs facilitated by Gregor Martius (Schubkraft).

The very mixed group of ~30 participants got a solid introduction to AI from a general perspective, and in a highly interactive session we all had a chance to vote our preferences, ask critical questions, and propose ways of impacting future development positively.

2.9. CH Open & BFH

In this evening event at the Bern University of Applied Sciences, we were shown by two expert engineers how to build a highly secure GenAI System based on popular Open Source tools. Projects like Ollama, Milvus, and Open WebUI were installed on premise in a live demo. For almost two hours we dived into details of software architecture, enterprise computing, certified supply chains, and Linux enhanced security features. I'm pretty sure most people were actually interested in the questions of data privacy and transparency.

Thanks Eric Lajoie and Holger Moenius for the interesting talk! Our video recording and the slide presentation is available.

3.9. Puzzle Breakfast

In this breakfast session, we packed into a coffee shop to hear of the latest developments in the AI ​​world from Puzzle, a local IT company. There was an intro to Agentic AI, the open-source protocols MCP and A2A. Solid business cases and ideas on how humans and AI can collaborate, what new ideas and ways of working are emerging. The two inspiring short presentations by Pascal Geronimi and Matthias Andenmatten gave insights into the trends, introduced by CEO Mark Waber.

Puzzle Breakfast - Neues aus der AI-Welt & ein konkreter Use Case

Künstliche Intelligenz verändert unsere Welt schneller denn je. In dieser Breakfast-Session teilen wir die neuesten Entwicklungen.

Puzzle ITCPascal Geronimi, Puzzle ITC

Thanks very much for your support of the Swiss {ai} Weeks, Puzzle! See also:

Ein Mate mit Pascal Geronimi – AI bei Puzzle

Wie verändert Künstliche Intelligenz unsere Arbeitswelt und wie stellen wir sicher, dass sie menschenzentriert und verantwortungsvoll eingesetzt wird?In unserer neuen Folge «Ein Mate mit…» gibt Pascal Geronimi spannende Einblicke in Chancen, Risiken und die Zukunft der KI. In unserer neuen Folge «Ein Mate mit…» gibt Pascal Geronimi spannende Einblicke in Chancen, Risiken und die Zukunft der KI.

Puzzle ITCFranziska Hashim

4.9. Polit-Forum Bern

As part of an inspiring and critical exhibition on "AI & Democracy", a Thursday night podium event dived into the topic of Fair and transparent AI – (how) does it work?

Artificial intelligence (AI) is increasingly being used to make processes and tasks faster and more efficient. However, this can jeopardize fundamental democratic rights. This is the case, for example, when AI participates in decisions regarding police work, social services, hiring procedures, or the granting of loans. Furthermore, the development and operation of AI models require enormous amounts of water and energy. Millions of workers are often trained under exploitative conditions. For years, AI ethicists have been demanding that fairness and transparency be guaranteed in the development and application of AI systems. But what does fair, transparent, and trustworthy AI look like in practice? ...

... What does (lack of) transparency mean for trust in AI? What are the risks if it's unclear how AI works? How and by whom should the production and use of AI systems be monitored and regulated? And what is needed to ensure that human rights, fairness, and sustainability are taken into account from the development to the application of AI technologies? (translated from German)

The panel discussion included:

• Rafael von Matt (moderator), Federal Palace Editor SRF
• Estelle Pannatier, Policy Manager at AlgorithmWatch CH
• Livia Walpen, Senior Policy Advisor, Federal Office of Communications
• Thilo Spinner, Research Engineer, IVIA Lab at ETH Zurich

I found the presentation and questions from the audience highly engaging, and am thankful to the Polit-Forum team for their extremely thorough and compelling take on the subject. During the apéro that followed the debate, I set up my laptop with PublicAI and several participants were able to chat or speak with Apertus. Here are the transcripts in German.

The current exhibition runs until mid-December.

KI und Demokratie: welche Zukunft wollen wir? - Polit-Forum Bern

Polit-Forum Bern

5.9. BFH

AI for SMEs: What Can Small Businesses Really Do with Artificial Intelligence? was a program that we designed to completent the activities with a perspective on the intersection of business and science departments. The University hosted us for an afternoon dedicated to demystifying AI for small and medium enterprises, taking place as part of the Swiss {ai} Weeks.

This event cuts through the hype to show you what's possible with AI in your business today – without the complexity, massive budgets, or technical expertise typically associated with AI implementation. ​What to expect: The event will begin with a keynote address by Jérôme Berthier, the CEO and founder of the Swiss generative AI company DeepLink, followed by a presentation on the Swiss AI initiative by Aram Davtyan of the University of Bern. These will be followed by interactive workshops in which you will gain hands-on experience of generative AI, prompt engineering, local AI models and business model innovation. Each session is designed to provide you with practical knowledge that you can implement immediately.

Thank you to Sarah Dégallier Rochat and team for a fantastic event!

👀 Last Fridays’s event in Bern – part of the national Swiss {ai} Weeks – tackled the question every small business owner is secretly asking: &quot;What can I actually do with AI that won&#39;t break my… | BFH Humane Digital Transformation

👀 Last Fridays’s event in Bern – part of the national Swiss {ai} Weeks – tackled the question every small business owner is secretly asking: “What can I actually do with AI that won’t break my budget or require a computer science degree?” 💡Jerome Berthier from Deeplink.ai opened with some refreshing honesty about AI being “a simple tool” rather than magic, while Aram Davtyan walked us through Switzerland’s AI Initiative and practical approaches. 💻 These insights were seamlessly built upon in the workshops. · Jerome Berthier also led a hands-on session diving into practical GenAI applications · Peter von Niederhäusern showed us generative AI tools that actually work · Thiemo Wambsganß had people rethinking smart prompting techniques · Stefan Hackstein explored something crucial: how to work with AI while keeping the cognitive effort that actually makes learning stick · Marcel Gygli gave practical insights to open-source alternatives What struck the most? People came with laptops and lots of curiosity and left with actual, implementable ideas. Thanks to everyone who showed up ready to learn and experiment! 👉🏼 Don’t want to miss events like this? Follow us to stay in the loop! Oleg Lavrovsky Sarah Dégallier Rochat

LinkedInBFH Humane Digital Transformation

6.9. Night of Research

While not an official {ai} Weeks event, I'd like to highlight a couple of impression of this annual occasion, during which dozens of laboratories from the University of Bern open doors and put up rather unique shows and activities for the general public. It was not a night to be missed!

Visit the official websites for more pictures, videos, links:

Rückblick 2025

Nacht der Forschung 2025

10.9. Polit-Forum

As part of the same exhibition described above, we discussed "AI in games and pop culture – a reality check?" in the format of a Democracy Bar.

Technological progress and associated visions of the future have long been explored in science fiction literature, films, games, and radio plays. Sometimes these are dazzling futures and images full of hope; at other times, bleak scenarios predominate, portraying technology as the oppressor of humanity and nature. In these imagined worlds, human life is determined by artificial intelligence, cyborgs—i.e., artificially enhanced living beings—and a highly technological environment. With the rapid spread of new AI applications today, the question arises: How far are we (still) from these worlds? (Translated from German)

Arno Görgen (historian of ideas and culture at HKB) and Eugen Pfister (historian of ideas and politics at HKB) addressed these and other questions, while everyone crowded around the bar to ask questions and join the discussion. This event is a collaboration with the CH Ludens project and the Bern University of the Arts (HKB).

15.9 F{ai}R

Yesterday we organized a first trade fair dedicated to AI at WORKSPACE & MORE - the new coworking space and bustling tech area of Liebefeld - the area of Bern where we are hosting our hackathon later this week. This busy Monday included a full program of events and showcases to launch a #siliconlovefield hashtag: a transliteration into English of the geographic name of Liebefeld (1st edition T-shirts are in stock!) as a local community initiative, coming soon to siliconlovefield.ch

AI in Action - Practical AI with chatbots, agents, and AI orchestration

Workshop with BEGASOFT AG

At the live showcase, you'll experience how an idea is transformed into functioning AI agents. First, we'll present the common foundation: We'll demonstrate our AI orchestration platform Brandbot , the use case, and our methodical approach—from evaluation to implementation and transformation.

MCP entdecken - Wie das Model Context Protocol KI erweitert

Workshop mit Unic

Experience how artificial intelligence transcends itself! In this exclusive workshop, Lukas, Niels, and Martin from Unic's Innovation Lab will demonstrate how the Model Context Protocol (MCP) connects AI systems with external tools, data sources, and real business power.

Hands-on AI experience: Learn prompting & Battle

Workshop with inovio GmbH

Experience generative KL live! In this practical workshop, we'll show you how to achieve fast results with clever prompts. After a short introduction, you'll practice in small groups before competing against each other in a prompting battle. Ideal for anyone who wants to understand KL in a playful way and apply it directly.

AI for SMEs: Maximize the use of ChatGPT & Co

Workshop with incratec GmbH

AI and SMEs are a perfect combination – AI can simplify many tasks and processes, saving money and time. Tobias Kluge, an AI expert from Bern, provides an overview and examples of how this can be achieved optimally: without expensive investments in AI tools, while observing data protection and security – and a concrete implementation plan.

Our exhibitors for the F{ai}R included:

Bevel ON | Empowering Displaced Professionalsto Build Careers

Our unique Programme approach equips displaced professionals with the tools necessary to restart their careers in a new environment.

Empowering Displaced Professionalsto Build Careers

High Five Academy

High Five Academy bietet maßgeschneiderte Kompetenzentwicklung – individuell auf deine Bedürfnisse und dein Wissen abgestimmt, damit du heute den Grundstein für deinen Erfolg von morgen legst.

inovio – Beratung für Digitale Transformation

Gemeinsam bringen wir bei inovio durch Technologie und Transformation Mehrwert in Dein Unternehmen – menschlich, verständlich und nachhaltig.

inovio GmbH

Home - Interprimis - Ihr Customer Experience Sourcing Hub

Transforming Dienstleistungen Trainings Academy Interprimis ist der erste Hub für intelligentes Sourcing von Customer Operations in der digitalen und hybriden Arbeitswelt. Wir sind spezialisiert auf individuelles Training von Experten sowie entwickeln wir hochanpassungsfähige Teams für B2C- und B2B-Kundenerfolg in spezifischen Branchen. Zukunftsfähig im Customer Operations Der Hub bietet Leistungen zur Förderung des Erfolgs aller Akteure […]

Interprimis - Ihr Customer Experience Sourcing Hub

KImpact - Verband für künstliche Intelligenz

Bei KImpact vernetzen sich KI-Enthusiasten und -Experten, um gemeinsam die Welt der Künstlichen Intelligenz zu erkunden. Mit exklusiven Workshops, spannenden Projekten und einer lebendigen Community bieten wir eine Plattform für Austausch und Lernen im Bereich KI. Werde auch du Mitglied!

Verband für künstliche Intelligenz

KIVY

Ihr Chatbot mit KI – Entwickelt in der Schweiz, trainiert mit Ihren Daten. Ihr persönlicher KI‑Chatbot – powered by Kivy, optimiert für Unternehmen in der Schweiz.

scb logo

netiva – Digitale Innovation für Unternehmen | IT-Services & Beratung

Die netiva unterstützt Unternehmen mit innovativen IT-Services & Beratung. Entdecken Sie zukunftssichere Lösungen für Ihre digitalen Herausforderungen.

logo der netiva-ag

AI

Changing IT for the better

Puzzle ITC

Digitalagentur in Zürich, Bern, München und Karlsruhe

Die 280 Expert:innen von Unic konzipieren, entwickeln und hosten für Sie Ihre digitale Plattform. Kontaktieren Sie uns für Ihr nächstes Webprojekt.

UnicUnic

Huge thanks to Selma and Sandro from WORKSPACE & MORE for their energetic support of the whole concept, and this massive day's program.

Attention listeners

During inovio's Prompt Battle and later in the Apéro, we got to hear a very exclusive live performance from rising electronic artist CyberGwen – a technologist performing on purely non-generative electronic synthesizers. Stop by our Hackathon at the very same location on Thursday night to experience her full show!

This is a big topic* in our community right now, and as Dribdat now also has recently had a first PR with purely generated code (yes, I'm very late to this party), I'm reaching out for opinion on both how to properly disclose - and support - working with AI at hackathons.

To maintain Dribdat I have relied Dependabot and FOSSA, along with various coding and continuous integration tools for years. Software fertilizer is hard to avoid these days! Codeberg's position seems clear that 'software is about humans', but someone is already working on Forgejo MCP support to allow Agents to interact with such sites. If there is an open source generator they endorse, I would certainly give it preference.

The main issue seems to be the loss of control. As long as a human reads and approves every line of every PR, it's a mitigatable concern. Try dealing with AI agents and their minions turning your backlog into an issue swamp. The other tricky one is the creative disenfranchising of millions of developers through coercive techniques on gamified platforms like StackOverflow and GitHub, the lack of foresight from OSI and other organisations in the legal area. We are only just beginning to use AI opt out - a.k.a. content exclusion controls, and that may feel too late for many people.

Very importantly we need to make clear which parts / commits are from a bot / agent, and take some time to understand the consequences** of code generation. I found the discussion at ghostty about disclosure quite relevant.

I can make sure that in our PRs and releases this is stated prominently. As you can see above, Jules and I recently added support for CC Signals, a prominent initiative in this direction. I'd also like to add a way to make GenAI commits visible and accounted for - see dribdat#426

Hackathons should respond to this. If we have an opportunity in public sprints to give some healthy advice to the next generation, and lead by example, I'd like to take it. So if you have any good examples or "betterHack" practices, I'm all ears.

The Git log doesn't lie.

* e.g. Nancy Law et al 2025 among others referenced in our hackathons library
** Petropakis et al 2024 etc.

Discuss this with us on GitHub:

Disclosing AI support at hackathons · dribdat · Discussion #29

This is a big topic* in our community right now, and as Dribdat now also has had a first PR with purely AI-generated code, I’m reaching out for opinion on both how to properly disclose - and suppor…

GitHubloleg