Maia Chess

Maia Chess is a research project out of the University of Toronto and Microsoft Research that takes a different approach to chess AI. Instead of training a neural network to find the best move, the team trained it to predict the move a human player would most likely make. The result is an engine that plays like a real person at a specific rating level, complete with the same kinds of mistakes, oversights, and tendencies you actually encounter in online games.

The team developed nine versions of Maia, each calibrated to a different Lichess rating level between 1100 and 1900. Every version was trained on millions of real human games at that rating range. The accuracy of human move prediction reaches up to 75% when the model is personalized to an individual player, meaning Maia can effectively model your specific chess style and tendencies.

You can play against Maia directly on the maiachess.com website, which offers realistic games, puzzle training, and opening drills against Maia models at your level. The bots are also available to play on Lichess under accounts like maia1, maia5, and maia9. Maia 2, published at NeurIPS 2024, extends the original work into a unified model covering the full rating spectrum with even higher prediction accuracy.