model_config.json

{
  "model_type": "BrainGemma3D",
  "language_model_dir": "language_model",
  "vision_model_dir": "vision_model",
  "depth": 2,
  "num_vision_tokens": 32
}