1e4_ai chess models beat MAIA-2 accuracy with human-like thinking time
Trained on 1B games, tiny 9M parameter models simulate human play across 800–2500+ ratings.
A developer (hazard02) has released 1e4_ai, a suite of transformer-based chess models designed to play like humans across a wide range of skill levels. The models are trained on nearly a billion games from Lichess, with separate models for each 100-point Elo bucket from ~800 to 2500+. Each bucket contains three dedicated 9-million-parameter models: one for move prediction, one for thinking time, and one for outcome probability (white win/draw/black win). The move models achieve better accuracy than MAIA-2 and are on par with MAIA-3, despite being much smaller. Training used an 8×H100 cluster for a mid-strength model, then fine-tuning on a single RTX 5090 for other rating ranges.
The project is notable for being the first to train on human thinking times in chess. The models factor in both player rating and remaining clock time, so a strong player under extreme time pressure has a lower win probability and blunders more often. The data pipeline, written in C++ with nanobind and PyTorch, was the most time-consuming part—pre-shuffling the dataset to keep GPUs busy during training. The code, training scripts, and model weights are fully open-source on GitHub, and a demo runs at 1e4.ai. This approach offers a more realistic AI opponent that mirrors human behavior, including time management and pressure-induced mistakes.
- Three models per rating bucket: move prediction, thinking time, and outcome probability (win/draw/loss).
- 9M parameters each, outperforming MAIA-2 and matching MAIA-3 on move accuracy.
- First model to integrate human thinking time data, causing realistic blunders under clock pressure.
Why It Matters
This open-source project enables realistic AI chess opponents that think and blunder like humans, useful for training and game analysis.