maia2-rapid

Maia 2 Rapid

This is the official Maia-2 checkpoint trained on human rapid games, packaged for inference with ChessNets. Maia-2 conditions its move predictions on both players' rating categories; it is not a calibrated estimate of an individual player's strength.

Use with ChessNets

from chessnets.pipelines.maia2 import Maia2Pipeline

pipeline = Maia2Pipeline.from_pretrained("shermansiu/maia2-rapid", device="cpu")
predictions = pipeline(
    "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1",
    elo_self=1500,
    elo_oppo=1600,
)

The model returns a legal-move probability distribution and a win-probability head. It uses the current position rather than a multi-position game history.

Checkpoint artifacts

• config.json: canonical ChessNets inference configuration
• model.safetensors: canonical inference tensors
• provenance.json: source hashes and exact conversion checks
• original/model.pt: original full PyTorch training checkpoint
• training/original_config.yaml: original upstream training configuration

The safetensors file contains only model_state_dict. The original optimizer state and training counters remain in original/model.pt. Conversion removes the DataParallel module. key prefix, then verifies every tensor and all three model outputs with exact torch.equal comparisons.

Original Maia-2 project: https://github.com/CSSLab/maia2

Original rapid checkpoint: https://drive.google.com/uc?id=1gbC1-c7c0EQOPPAVpGWubezeEW8grVwc

Original training config: https://drive.google.com/uc?id=1GQTskYMVMubNwZH2Bi6AmevI15CS6gk0

Intended use and limitations

This checkpoint is intended for research into human move prediction and skill-conditioned chess behavior. Its predictions reflect patterns in the training population and time-control domain. They are not guarantees of playing strength, fairness, instructional quality, or suitability for high-stakes decisions.

License

The upstream Maia-2 code and published model are distributed under the MIT License. ChessNets is a separate software package.

Citation

@inproceedings{tang2024maia,
  title={Maia-2: A Unified Model for Human-{AI} Alignment in Chess},
  author={Zhenwei Tang and Difan Jiao and Reid McIlroy-Young and Jon Kleinberg and Siddhartha Sen and Ashton Anderson},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems},
  year={2024},
  url={https://openreview.net/forum?id=XWlkhRn14K}
}