Backbone Models
Code repair and GUI interaction use task-specific checkpoints with upstream backbone models. Download those backbones and keep them in a shared model directory outside this collection.
| Model and download | Used by | Role |
|---|---|---|
| ReasonFlux-Coder-7B | Code repair, GUI interaction | Program generation and action proposal. |
| Qwen2.5-Coder-3B-Instruct | Code repair, GUI interaction | Encoding backbone for the JEPA world models. |
Download
After installing the code-repair or GUI requirements, use the hf command to download its required models. These commands select the versions used with the released checkpoints:
export RSI_BASE_MODELS_DIR="$HOME/.cache/rsi_anything/base_models"
mkdir -p "$RSI_BASE_MODELS_DIR"
hf download Gen-Verse/ReasonFlux-Coder-7B \
--revision 6860a2fd222bd2e18e8a5b62a9c0887d2ba70c36 \
--local-dir "$RSI_BASE_MODELS_DIR/ReasonFlux-Coder-7B"
hf download Qwen/Qwen2.5-Coder-3B-Instruct \
--revision 488639f1ff808d1d3d0ba301aef8c11461451ec5 \
--local-dir "$RSI_BASE_MODELS_DIR/Qwen2.5-Coder-3B-Instruct"
Code repair and GUI interaction use both models. Laboratory planning loads its local JEPA world models directly. Reaction optimization loads its local ReactionT5 + JEPA yield world model directly. These entries support CPU inference; the reaction-yield model also supports CUDA.
Model paths
Entry points read RSI_BASE_MODELS_DIR, which defaults to ~/.cache/rsi_anything/base_models. An existing model directory can be used by setting this variable to its parent folder. Keep the weights, configuration, and tokenizer together in each model folder:
$RSI_BASE_MODELS_DIR/
├── ReasonFlux-Coder-7B/
└── Qwen2.5-Coder-3B-Instruct/
Domain adapters and predictive heads remain in each task's checkpoints/ folder and are loaded automatically. The _base_models/ folder in this collection contains this guide and the original model license and attribution materials. See the third-party notices for model terms and MODEL_INDEX.json for the resource mapping.