LayaChess

A general-purpose AI decision model taught to play chess. Laya (Convai Innovations, ModernBERT-large, 421M, Apache 2.0) was never trained on chess. LayaChess fine-tunes it on DeepMind's ChessBench and wraps it in a Leela-style search engine.

• Play it online: huggingface.co/spaces/datafreak/laya-chess
• Write-up: LayaChess: teaching a System 1 decision model to play chess
• Demo video: youtube.com/watch?v=bPpAlWArs7E
• Model: datafreak/laya-chess

Results (v2, 2 Kaggle sessions on 2× T4, 2.05M training examples)

	Base Laya	LayaChess v2
Picks Stockfish's best move (300 held-out positions)	6%	27%
Win-chance error (percentage points)	28.6	8.2
Opening move as White	b4	d4
Mate-in-one (Scholar's mate, Qxf7#)	missed (prefers d3)	found, 94% win chance

Strength in games: still below Stockfish at its weakest official setting (UCI_Elo 1320) on a laptop. A Lichess bot rating is next.

How it works

• Question per move (DeepMind's action-value recipe, asked in Laya's own format): for each legal move Laya gets a score question such as "white plays Nf3 (knight g1-f3). Win chance for white?" with 10 win-chance levels; the board is given as piece lists. The answer's expected value is the move's win chance.
• Engine: MCTS / PUCT (AlphaZero/Leela style). Laya's move scores become priors and values; mate, stalemate and draws come from the rules. UCI front end, browser board, match runner. See engine/README.md.
• v3 (in progress): read the board once and score every move in one pass with Laya's encoder + a move-query head (~35× faster search), warm-started from v2.

Run it on your own machine

You need Python 3.10+ and about 2 GB of disk for the model. A GPU helps but isn't required; on a Mac it uses Apple's GPU (MPS) automatically.

1. Get the code

   git clone https://github.com/devroopsaha744/LayaChess.git
   cd LayaChess/engine

2. Create an environment and install (with uv, or plain pip below)

   uv venv --python 3.12 .venv && source .venv/bin/activate
   uv pip install -r requirements.txt

   Without uv: python3 -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
3. Optional: install Stockfish, to see Stockfish's preferred move next to Laya's

   brew install stockfish          # macOS
   sudo apt install stockfish      # Debian / Ubuntu

4. Play

   python -m laya_chess.play       # opens the board at http://localhost:8000

   The first start downloads the model from Hugging Face (about 850 MB), so give it a minute.

Pick your colour and how long Laya thinks: Instantly is the network alone, 3 to 30 seconds adds the tree search. More commands (UCI engine for chess GUIs, matches against Stockfish) are in engine/README.md.

Files

Path	What
notebooks/laya_chess_finetune.ipynb	v2 fine-tuning (Kaggle 2× T4, fp16, checkpoints to HF, resume)
notebooks/laya_chess_v3_kaggle.ipynb	v3 training (one pass per position)
notebooks/laya_vs_stockfish_kaggle.ipynb	matches vs Stockfish on Kaggle
notebooks/chessbench_download.ipynb	ChessBench files → Kaggle dataset
notebooks/build_*.py	generate the notebooks above
engine/	model loading, MCTS, UCI, browser board, matches, tests
space/	the Hugging Face Space (Gradio + ZeroGPU); ./deploy.sh uploads it
docs/	learning curve, plan, post draft

Credits

Laya by Convai Innovations (Apache 2.0) · ChessBench / searchless chess by Google DeepMind · Stockfish · python-chess