Angelotxx271/chess-transformer-demo

Chess Transformer

Play against a 39M-parameter dual-head transformer with cross-temporal attention over 8 board states. Trained from scratch on Lichess games as part of an Individual Project.

Training: ~5 days of supervised pretraining + ~3 days of PPO self-play on a single RTX 5070.

Code on GitHub | Model card

Free CPU Spaces are slow — set the MCTS slider to 0 for snappy moves; 200 sims will take ~10 s per move on CPU.