9M-parameter chess transformers match MAIA-3, first to model thinking time
A researcher trained transformer models on 1 billion Lichess games to replicate human chess play across rating bands from 800 to 2500+, including separate networks for move selection, thinking time, and win probability.
A researcher trained transformer-based chess models on nearly a year of Lichess data—roughly 1 billion games—to replicate human play across skill levels from 800 to 2500+ Elo. Each 100-point rating bucket gets three separate 9-million-parameter networks: one for move selection, one for thinking time, and one for win/draw/loss probability. The move models match MAIA-3 accuracy, the current benchmark for human-like chess AI, despite being significantly smaller.
The project breaks new ground by modeling thinking time—reportedly the first attempt to train on how long humans deliberate before moving. The models condition on player ratings and clock times, blundering more under time pressure and adjusting win probabilities when a stronger player faces extreme time constraints. At higher ratings the models spot short tactical patterns but struggle with deep calculation, a function of network size.
Training started on an 8×H100 cluster for a mid-strength baseline, then fine-tuned rating-specific variants on a local RTX 5090. The researcher spent most development time on the data pipeline—a C++ nanobind interface feeding PyTorch—pre-shuffling the dataset to keep GPU utilization high and avoid I/O bottlenecks. Code, weights, and a live demo are available on GitHub and at 1e4.ai.