This directory contains AI models for EMBODIOS. Models are downloaded on-demand and not stored in git.

Download the default test model (TinyLlama):

./scripts/download-models.sh

Download a specific model:

./scripts/download-models.sh phi-2

All models are tracked in manifest.json with:

• Exact version and revision
• SHA256 checksums for integrity
• Download URLs
• File sizes

This ensures reproducible testing across different environments.

Pull models using the CLI:

# Pull from HuggingFace
embodi pull TinyLlama/TinyLlama-1.1B-Chat-v1.0

# Pull with quantization
embodi pull microsoft/phi-2 --quantize 4

# Pull direct URL
embodi pull https://huggingface.co/.../model.gguf

models/
├── manifest.json      # Model registry with versions
├── README.md         # This file
├── .gitkeep          # Ensures directory exists in git
└── tinyllama/        # Downloaded models (git-ignored)
    ├── tinyllama-1.1b-chat-v1.0.Q4_K_M.gguf
    ├── .version      # Version tracking
    └── .revision     # Git revision tracking

1. Update manifest.json with model details:

   • Version and revision
   • Download URL
   • SHA256 checksum
   • File size

2. Run download script:

   ./scripts/download-models.sh your-model

Run tests that use real models:

# Download model first
./scripts/download-models.sh

# Run tests
pytest tests/integration/test_real_model_inference.py

Run interactive demo:

./scripts/testing/test_interactive.sh

• Models are NOT stored in git (see .gitignore)
• Always use the download script to ensure correct versions
• The manifest ensures everyone uses the same model versions
• Models are cached locally after first download