UniMaia

UniMaia is a 3,166,691,262-parameter prompt-conditioned chess policy model from UniMaia: Steering Chess Policies with Language for Human-like Play. It combines an Lc0 attention-body policy network with a LoRA-adapted ChessGPT text encoder and ControlNet-style conditioning. This variant returns prompt-conditioned chess move probabilities.

Usage

Install chessnets with its UniMaia inference dependencies, then provide a PGN prefix and a natural-language description of the desired player or behavior:

from chessnets import AutoPipeline

pipeline = AutoPipeline.from_pretrained("shermansiu/unimaia", device="cuda")
predictions = pipeline.make_prediction(
    pgns="1. e4 e5 2. Nf3 Nc6 *",
    prompts="A cautious club player is playing White.",
)
print(predictions[0].best_move_san)

PGNs must describe legal chess positions. Prompts can describe attributes such as player strength, opening preferences, or playing style. The conditioning is probabilistic: it steers the policy rather than enforcing a hard constraint.

Intended use

This checkpoint is intended for research and chess applications involving prompt-conditioned move prediction, human-like chess modeling, controllable chess agents, and analysis of language-conditioned policies through chessnets. It is not a general-purpose language model or a guarantee of engine strength.

Limitations

• Training prompts were produced primarily from synthetic natural-language templates over game metadata. The model has not been trained on the full diversity of natural human instructions.
• The learned policy is sensitive to prompt wording, particularly early prompt tokens. Attention-sink behavior can make superficially similar phrasings produce different move distributions.
• Out-of-distribution wording, attributes, languages, chess variants, or contradictory instructions may be ignored or interpreted unpredictably.
• Prompt conditioning is statistical and does not guarantee compliance with a requested opening, strength, identity, or style.
• Results inherit limitations and biases from the Lichess training data, the base Lc0 network, ChessGPT, and the prompt-generation process.

Training and evaluation

The model was trained on metadata-augmented Lichess games using generated prompt templates. Architecture, training, and benchmark details are reported in the paper. Exact source configuration, checkpoint provenance, and file hashes are included in provenance.json and artifact_manifest.json. This release corresponds to controlnet_onephase_full_training_run_v1.

License and upstream models

The model artifacts in this repository are released under GPL-3.0. This model is derived from an Lc0 GPL-3.0 model and the ChessGPT Apache-2.0 text model. This model-artifact license does not relicense the separately distributed UniMaia, chessnets, or chesseval software; consult those repositories for their code and third-party-file licensing.

Citation

The paper is currently under review. Please use the archival EMNLP Findings citation if the paper is accepted and that version becomes available; until then, cite the arXiv version:

@article{siu2026unimaia,
  title={UniMaia: Steering Chess Policies with Language for Human-like Play},
  author={Siu, Sherman and Istead, Lesley},
  journal={arXiv preprint arXiv:2605.27767},
  year={2026},
  doi={10.48550/arXiv.2605.27767}
}

• arXiv: https://arxiv.org/abs/2605.27767
• Review page: https://openreview.net/forum?id=ARQY5L1frw