Promote to
Maia-3 is a neural network from the University of Toronto's CSSLab trained to predict the move a human at a given rating would play — not the best move. This app runs its 5M-parameter version entirely in your browser, so there is no server.
Each turn the model returns a probability for every legal move. The opponent samples from that distribution, so it makes the mistakes players at that rating make, in roughly the proportions they make them.
Playstyles tilt Maia's probabilities toward measurable move features (checks, captures, advancing, safety, castling). The style report compares what the opponent actually played with what plain Maia would have done in the same positions, and shows the cost in human-likeness (KL divergence).
Limits: style weights are hand-tuned, not learned from real players; the rating slider sets Maia's conditioning, it is not a calibrated playing strength; the clock is for you — Maia-3 was not trained on clock times, so it does not play faster or worse under time pressure the way humans do.
Maia-3 © CSSLab, AGPL-3.0 · TypeScript port: maia3-ts (AGPL-3.0) · Rules: chess.js (BSD-2). This app is AGPL-3.0; download the full source from the header link.