Maia Chess

Maia Chess is a research project from the University of Toronto and Microsoft Research that rethinks what a chess AI should be designed to do. Most chess engines are built to find the strongest possible move. Maia was built to predict the move a human player would actually make, including the mistakes.

The team trained nine separate neural network models, each calibrated to a specific Elo range from 1100 to 1900. Each model was trained on 12 million real Lichess games played at that rating level, rather than games between AIs or top grandmasters. The result is a set of bots that feel genuinely human: they blunder in ways your peers blunder, they miss the same kinds of tactics you miss, and they respond to positions the way someone at your level actually would.

This makes Maia unusually useful for practice. Instead of playing an engine that either crushes you or plays randomly, you can play against a version of Maia calibrated to your own rating, getting realistic opposition without the frustration of facing inhuman perfection or trivially weak moves.

Maia 2, presented at NeurIPS 2024, extended the original work with a unified 23.3M parameter model covering the full rating spectrum. It achieves better human move prediction accuracy while requiring far fewer parameters than the original nine-model setup.

Maia is playable for free on maiachess.com and directly on Lichess as a bot opponent. The full codebase is open-source and available on GitHub through the CSSLab at the University of Toronto.