Otter chess AI predicts human moves with 55% accuracy, beats Maia 2

Otter is a time-aware, history-conditioned chess AI that reframes move prediction as a sequential process, not a single-position snapshot. Built by Tarun Kumar S, the 15.3M-parameter model uses two conditioning signals: a move history encoder that tracks the last 20 moves to capture opening preferences and positional drift, and a time control module that adjusts predictions based on clock pressure. This dual approach lets Otter behave more like a human under time stress, where rapid or flagging decisions often differ from optimal play.

Trained on 6.1 billion positions from 117 million Lichess rapid games over 30 days on a single T4 GPU, Otter achieves 55.23% top-1 and 90.95% top-5 move-prediction accuracy. It surpasses the prior state-of-the-art Maia 2 model while using far fewer parameters and less training data. Accuracy peaks at 57.38% in the 1900-1999 Elo bracket and holds across 11 Elo brackets from under 1100 to 2000+. The results demonstrate that modeling chess as a time-aware, sequential human activity yields more human-accurate predictions than position-only models. Code, trained models, and training logs are publicly released.