ChessMimic: Per-Rating Transformer Models for Human Move, Clock, and Outcome Prediction in Online Blitz Chess

Researchers introduce ChessMimic, a system of three transformer models that predict human chess moves, thinking time, and game outcomes in online blitz chess with rating-specific calibration. The models outperform existing systems like Maia across multiple performance metrics while using significantly fewer parameters, with code and weights publicly released.

ChessMimic represents a meaningful advance in human behavior prediction within constrained domains, demonstrating that specialized, smaller models can outperform larger generalist systems when properly calibrated. The research departs from traditional scaling approaches by training separate model instances per 100-Elo rating band, sacrificing parameter efficiency for sharper skill-level discrimination. This architectural choice reflects a broader principle in machine learning: domain-specific optimization often beats raw model size when predicting nuanced human behavior.

The system's three-model approach addresses distinct prediction tasks—move selection, clock management, and outcome assessment—each influenced by different contextual factors. Move prediction achieves state-of-the-art results across all rating bands compared to Maia-2, while the outcome model incorporates player ratings and time dynamics to reach 0.78 AUC. The clock model shows functional but non-optimal performance at predicting thinking times, with residual accuracy gaps concentrated in position-specific buckets rather than overall calibration.

For the AI research community, ChessMimic validates that efficient, interpretable models can compete with larger alternatives when engineered thoughtfully. The public release of code, per-band weights, and C++ infrastructure reduces barriers to reproducibility and extension. The approach has potential applications beyond chess—any domain involving skill-stratified human decision-making under time pressure could benefit from similar per-tier modeling strategies. The work underscores that competitive AI systems need not maximize parameters; instead, architectural choices that respect domain structure and human heterogeneity drive performance gains.

• →ChessMimic outperforms larger models like Maia-2 while using 9M parameters by using rating-band-specific calibration
• →The system trains three separate transformers for move, clock, and outcome prediction conditioned on position and player context
• →Public release of code, weights, and C++ pipeline democratizes access to the models and methodology
• →Rating-stratified modeling demonstrates that domain-specific optimization beats generalist scaling in human behavior prediction
• →Outcome model achieves 0.78 AUC by incorporating player ratings and clock time alongside board position