Docker Model Runner (DMR) makes it easy to manage, run, and deploy AI models using Docker. Designed for developers, Docker Model Runner streamlines the process of pulling, running, and serving large language models (LLMs) and other AI models directly from Docker Hub or any OCI-compliant registry.

This package supports the Docker Model Runner in Docker Desktop and Docker Engine.

For macOS and Windows, install Docker Desktop:

https://docs.docker.com/desktop/

Docker Model Runner is included in Docker Desktop.

For Linux, install Docker Engine from the official Docker repository:

curl -fsSL https://get.docker.com | sudo bash
sudo usermod -aG docker $USER # give user permission to access docker daemon, relogin to take effect

Docker Model Runner is included in Docker Engine when installed from Docker's official repositories.

To verify that Docker Model Runner is available:

# Check if the Docker CLI plugin is available
docker model --help

# Check Docker version
docker version

# Check Docker Model Runner version
docker model version

# Run a model to test the full setup
docker model run ai/gemma3 "Hello"

If docker model is not available, see the troubleshooting section below.

If you encounter errors like Package 'docker-model-plugin' has no installation candidate or docker model command is not found:

1. Check your Docker installation source:

   # Check Docker version
   docker version
   
   # Check Docker Model Runner version
   docker model version

   Look for the source in the output. If it shows a package from your distro, you'll need to reinstall from Docker's official repositories.

2. Remove the distro version and install from Docker's official repository:

   # Remove distro version (Ubuntu/Debian example)
   sudo apt-get purge docker docker.io containerd runc
   
   # Install from Docker's official repository
   curl -fsSL https://get.docker.com | sudo bash
   
   # Verify Docker Model Runner is available
   docker model --help

3. For NVIDIA DGX systems: If Docker came pre-installed, verify it's from Docker's official repositories. If not, follow the reinstallation steps above.

For more details refer to:

https://docs.docker.com/ai/model-runner/get-started/

Before building from source, ensure you have the following installed:

• Go 1.25+ - Required for building both model-runner and model-cli
• Git - For cloning repositories
• Make - For using the provided Makefiles
• Docker (optional) - For building and running containerized versions
• CGO dependencies - Required for model-runner's GPU support:
  • On macOS: Xcode Command Line Tools (xcode-select --install)
  • On Linux: gcc/g++ and development headers
  • On Windows: MinGW-w64 or Visual Studio Build Tools

After cloning, a single make builds everything — the server, CLI plugin, and a dmr convenience wrapper:

make

dmr starts the server on a free port, waits for it to be ready, runs your CLI command, then shuts the server down:

./dmr run ai/smollm2 "Hello, how are you?"
./dmr ls
./dmr run qwen3:0.6B-Q4_0 tell me today's news

These components can also be built, run, and tested separately using the Makefile.

Note: We use port 13434 in these examples to avoid conflicts with Docker Desktop's built-in Model Runner, which typically runs on port 12434.

1. Start model-runner in one terminal:

MODEL_RUNNER_PORT=13434 ./model-runner

2. Use model-cli in another terminal:

# List available models
MODEL_RUNNER_HOST=http://localhost:13434 ./cmd/cli/model-cli list

# Pull and run a model
MODEL_RUNNER_HOST=http://localhost:13434 ./cmd/cli/model-cli run ai/smollm2 "Hello, how are you?"

1. Build and run model-runner in Docker:

cd model-runner
make docker-build
make docker-run PORT=13434 MODELS_PATH=/path/to/models

2. Connect with model-cli:

cd cmd/cli
MODEL_RUNNER_HOST=http://localhost:13434 ./model-cli list

• Model Runner Documentation
• Model CLI README
• Model Specification
• Community Slack Channel

This project includes a Makefile to simplify common development tasks. Docker targets require Docker Desktop >= 4.41.0. Run make help for a full list, but the key targets are:

• build - Build the Go application
• build-cli - Build the CLI (docker-model plugin)
• install-cli - Build and install the CLI as a Docker plugin
• docs - Generate CLI documentation
• run - Run the application locally
• clean - Clean build artifacts
• test - Run tests
• validate-all - Run all CI validations locally (lint, test, shellcheck, go mod tidy)
• lint - Run Go linting with golangci-lint
• validate - Run shellcheck validation on shell scripts
• integration-tests - Run integration tests (requires Docker)
• docker-build - Build the Docker image for current platform
• docker-run - Run the application in a Docker container with TCP port access and mounted model storage
• help - Show all available targets and configuration options

The application can be run in Docker with the following features enabled by default:

• TCP port access (default port 8080)
• Persistent model storage in a local models directory

# Run with default settings
make docker-run

# Customize port and model storage location
make docker-run PORT=3000 MODELS_PATH=/path/to/your/models

This will:

• Create a models directory in your current working directory (or use the specified path)
• Mount this directory into the container
• Start the service on port 8080 (or the specified port)
• All models downloaded will be stored in the host's models directory and will persist between container runs

The Docker image includes the llama.cpp server binary from the docker/docker-model-backend-llamacpp image. You can specify the version of the image to use by setting the LLAMA_SERVER_VERSION variable. Additionally, you can configure the target OS, architecture, and acceleration type:

# Build with a specific llama.cpp server version
make docker-build LLAMA_SERVER_VERSION=v0.0.4

# Specify all parameters
make docker-build LLAMA_SERVER_VERSION=v0.0.4 LLAMA_SERVER_VARIANT=cpu

Default values:

• LLAMA_SERVER_VERSION: latest
• LLAMA_SERVER_VARIANT: cpu

Available variants:

• cpu: CPU-optimized version
• cuda: CUDA-accelerated version for NVIDIA GPUs
• rocm: ROCm-accelerated version for AMD GPUs
• musa: MUSA-accelerated version for MTHREADS GPUs
• cann: CANN-accelerated version for Ascend NPUs

The binary path in the image follows this pattern: /com.docker.llama-server.native.linux.${LLAMA_SERVER_VARIANT}.${TARGETARCH}

The Docker image also supports vLLM as an alternative inference backend.

To build a Docker image with vLLM support:

# Build with default settings (vLLM 0.12.0)
make docker-build DOCKER_TARGET=final-vllm BASE_IMAGE=nvidia/cuda:13.0.2-runtime-ubuntu24.04 LLAMA_SERVER_VARIANT=cuda

# Build for specific architecture
docker buildx build \
  --platform linux/amd64 \
  --target final-vllm \
  --build-arg BASE_IMAGE=nvidia/cuda:13.0.2-runtime-ubuntu24.04 \
  --build-arg LLAMA_SERVER_VARIANT=cuda \
  --build-arg VLLM_VERSION=0.12.0 \
  -t docker/model-runner:vllm .

The vLLM variant supports the following build arguments:

• VLLM_VERSION: The vLLM version to install (default: 0.12.0)
• VLLM_CUDA_VERSION: The CUDA version suffix for the wheel (default: cu130)
• VLLM_PYTHON_TAG: The Python compatibility tag (default: cp38-abi3, compatible with Python 3.8+)

The vLLM variant supports both x86_64 (amd64) and aarch64 (arm64) architectures. The build process automatically selects the appropriate prebuilt wheel:

• linux/amd64: Uses manylinux1_x86_64 wheels
• linux/arm64: Uses manylinux2014_aarch64 wheels

To build for multiple architectures:

docker buildx build \
  --platform linux/amd64,linux/arm64 \
  --target final-vllm \
  --build-arg BASE_IMAGE=nvidia/cuda:12.9.0-runtime-ubuntu24.04 \
  --build-arg LLAMA_SERVER_VARIANT=cuda \
  -t docker/model-runner:vllm .

To update to a new vLLM version:

docker buildx build \
  --target final-vllm \
  --build-arg VLLM_VERSION=0.11.1 \
  -t docker/model-runner:vllm-0.11.1 .

The vLLM wheels are sourced from the official vLLM GitHub Releases at https://github.com/vllm-project/vllm/releases, which provides prebuilt wheels for each release version.

The Model Runner exposes a REST API that can be accessed via TCP port. You can interact with it using curl commands.

When running with docker-run, you can use regular HTTP requests:

# List all available models
curl http://localhost:8080/models

# Create a new model
curl http://localhost:8080/models/create -X POST -d '{"from": "ai/smollm2"}'

# Get information about a specific model
curl http://localhost:8080/models/ai/smollm2

# Chat with a model
curl http://localhost:8080/engines/llama.cpp/v1/chat/completions -X POST -d '{
  "model": "ai/smollm2",
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Hello, how are you?"}
  ]
}'

# Delete a model
curl http://localhost:8080/models/ai/smollm2 -X DELETE

# Get metrics
curl http://localhost:8080/metrics

The response will contain the model's reply:

{
  "id": "chat-12345",
  "object": "chat.completion",
  "created": 1682456789,
  "model": "ai/smollm2",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "I'm doing well, thank you for asking! How can I assist you today?"
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 24,
    "completion_tokens": 16,
    "total_tokens": 40
  }
}

• Automatic GPU Detection: Automatically configures NVIDIA GPU support if available
• Persistent Caching: Models are cached in ~/.cache/nim (or $LOCAL_NIM_CACHE if set)
• Interactive Chat: Supports both single prompt and interactive chat modes
• Container Reuse: Existing NIM containers are reused across runs

Single prompt:

docker model run nvcr.io/nim/google/gemma-3-1b-it:latest "Explain quantum computing"

Interactive chat:

docker model run nvcr.io/nim/google/gemma-3-1b-it:latest
> Tell me a joke
...
> /bye

• NGC_API_KEY: Set this environment variable to authenticate with NVIDIA's services
• LOCAL_NIM_CACHE: Override the default cache location (default: ~/.cache/nim)

NIM containers:

• Run on port 8000 (localhost only)
• Use 16GB shared memory by default
• Mount ~/.cache/nim for model caching
• Support NVIDIA GPU acceleration when available

The Model Runner exposes the metrics endpoint of llama.cpp server at the /metrics endpoint. This allows you to monitor model performance, request statistics, and resource usage.

# Get metrics in Prometheus format
curl http://localhost:8080/metrics

• Enable metrics (default): Metrics are enabled by default
• Disable metrics: Set DISABLE_METRICS=1 environment variable
• Monitoring integration: Add the endpoint to your Prometheus configuration

Check METRICS.md for more details.

Experimental support for running in Kubernetes is available in the form of a Helm chart and static YAML.

If you are interested in a specific Kubernetes use-case, please start a discussion on the issue tracker.

dmrlet is a purpose-built container orchestrator for AI inference workloads. Unlike Kubernetes, it focuses exclusively on running stateless inference containers with zero configuration overhead. Multi-GPU mapping "just works" without YAML, device plugins, or node selectors.

# Build the dmrlet binary
go build -o dmrlet ./cmd/dmrlet

# Verify it works
./dmrlet --help

Serve a model:

# Auto-detect backend and GPUs
dmrlet serve gemma3

For general questions and discussion, please use Docker Model Runner's Slack channel.

For discussions about issues/bugs and features, you can use GitHub Issues and Pull requests.