# clawd-local-ai-chess

Two large local LLMs playing chess against each other on a single Mac, refereed
by `python-chess`. **Step-3.7-Flash** (198B MoE, ~11B active) plays White against
**DeepSeek-V4-Flash** as Black, both running locally via `llama.cpp` with Metal.

Neither model fits alongside the other in 128GB of RAM, so the arena loads each
engine **one-shot per move** (load → get move → exit → free RAM). `python-chess`
is the sole source of truth: it builds the prompt, enforces legality, and decides
the result. The models never trust each other's moves.

## Files

- `chess/arena.py` — the referee / match runner
- `chess/game.pgn` — last game in PGN
- `chess/game.log` — full move-by-move log of the last game
- `chat.sh` — interactive chat with Step-3.7-Flash via `llama-cli`
- `models/chat_template.jinja` — Step-3.7-Flash chat template
- `models/chat_template_nothink.jinja` — variant with thinking disabled

## Last result

```
Event: Local LLM Heavyweight Match
White: Step-3.7-Flash   Black: DeepSeek-V4-Flash
Result: 1-0 (White wins)
```

## Setup (not committed here)

The model weights (~98GB of GGUF shards) and the `stepfun-ai/llama.cpp` fork are
intentionally **not** in this repo — see `.gitignore`. To reproduce:

1. Build the StepFun llama.cpp fork: <https://github.com/stepfun-ai/llama.cpp>
2. Download the GGUF weights for Step-3.7-Flash and DeepSeek-V4-Flash.
3. Fix up the absolute paths in `chess/arena.py` and `chat.sh`.
4. Run the arena:

   ```bash
   uv run --with chess python chess/arena.py
   ```