Maia-3: Open-Source Human-Like Chess AI Hits 57.1% Move Prediction Accuracy — Released on Lichess
Lichess / Hacker News·low signal
University of Toronto's CSSLab released Maia-3 on May 24, featuring 'Chessformer' — a transformer-based architecture treating 64 board squares as tokens with Geometric Attention Bias. Move prediction accuracy jumped to 57.1% (from Maia-2's 52.0%), modeling players across 600-2600 rating. Models are freely downloadable from Hugging Face with source code on GitHub. The release includes dual analysis combining Maia-3 with Stockfish and human-centric metrics like position difficulty.