ChessStep uses Maia3 5M by the University of Toronto CSSLab.
Project: https://github.com/CSSLab/maia3
Pinned source: 1e13597c42d4858b7cfd7cfdae01e297263364b2
Model: https://huggingface.co/UofTCSSLab/Maia3-5M
Pinned model revision: b6559de2398d7140b985f28fd2c19fb5e47ddabe

Maia3 and this standalone browser adapter: GNU AGPL version 3 (LICENSE.txt).
Modifications: policy-only ONNX export; equivalent RMSNorm decomposition;
JavaScript position encoding and legal policy sampling; Web Worker runtime;
Python CPU inference service (Cloudflare Container).
Original float32 model weights are unchanged; no quantization or fine-tuning.
Corresponding source, original checkpoint and rebuild instructions:
https://chessstep.com/maia/maia3-source.tar.gz

ONNX Runtime Web 1.30.0: MIT (ONNXRUNTIME-LICENSE.txt).
Runtime source: https://github.com/microsoft/onnxruntime/tree/v1.30.0

Reference: Monroe et al., Chessformer: A Unified Architecture for Chess Modeling,
2026. https://arxiv.org/abs/2605.19091

The reference rating conditions predicted human-style moves; it is not a
certified or measured playing rating. No affiliation or endorsement is implied.