ChessIQ Maia 3 browser engine
Integration date: 2026-09-28

Maia 3 (Chessformer), 5M parameters, by Daniel Monroe, George Eilender,
Philip Chalmers, Zhenwei Tang, Ashton Anderson and CSSLab contributors.
Official source: https://github.com/CSSLab/maia3
Revision: 1e13597c42d4858b7cfd7cfdae01e297263364b2
Upstream source is included in maia3-upstream-source.tar.gz.
Official original weights: https://huggingface.co/UofTCSSLab/Maia3-5M
Paper: https://arxiv.org/abs/2605.19091
License: GNU AGPL v3; see LICENSE.txt. No warranty.

ONNX weights: maia3_5m.onnx, distributed by Cem Unuvar in maia3-js 0.2.0.
SHA-256: f7bd01b3e41dceb5e93eb9f4c61e9c4db8b35411effd1d8c984b006cd74bb804
https://huggingface.co/cemoss17/maia3-onnx
The converter's MIT package notice does not replace the upstream Maia license.
The unchanged converted parameters and graph are provided locally; original
editable PyTorch parameters and architecture are at the official sources above.

maia3-js 0.2.0, copyright 2026 Cem Unuvar, MIT (maia3-js-LICENSE.txt).
Unminified modules are in lib/. Only module import paths and source-map comments
were changed for self-hosting. Package: https://www.npmjs.com/package/maia3-js
ONNX Runtime Web 1.30.0, copyright Microsoft Corporation, MIT.
See ort/LICENSE.txt and ort/ThirdPartyNotices.txt.
Runtime source: https://github.com/microsoft/onnxruntime/tree/v1.30.0
chess.js 1.4.0, BSD-2-Clause. See /vendor/chess-LICENSE.txt.

The browser worker and UCI adapter are ChessIQ modifications under
AGPL-3.0-or-later. They run separately from the interface and communicate through
UCI text. No model inference or board position is sent to an external service.
The 5M model uses a single WebAssembly thread; it is loaded only when selected.

Source: /engines/maia-3/chessiq-maia-source.tar.gz
Includes the adapter, unminified inference modules, upstream engine source,
chess.js, licenses, model metadata and the vendoring script. Reproduce the
third-party assets with Node 22+ by running scripts/vendor-maia.mjs from the
extracted root. The script verifies npm integrity and the pinned model hash.
No changes were made to the model parameters. Maia models human move choices;
ratings and win/draw/loss predictions are not Stockfish strength or evaluation.