maia3-onnx-int32

Maia3 ONNX Models (int32 ELO inputs)

Maia3 neural network models with int32 ELO inputs for Safari/WebKit compatibility.

Original models use int64 for self_elo and oppo_elo inputs, which Safari's WebAssembly doesn't support (no BigInt64Array). These models add Cast nodes (int32 → int64) at the input, allowing int32 tensors while keeping internal computation unchanged.

Models

Tensor Specs

• Input: tokens — [batch, 64, 96] float32
• Input: self_elo — [1] int32 (player ELO, 0-5000)
• Input: oppo_elo — [1] int32 (opponent ELO, 0-5000)
• Output: logits_move — [batch, 4352] float32 (policy logits)
• Output: logits_value — [batch, 3] float32 (loss, draw, win)

License

AGPL-3.0. The authoritative source for these weights is the official CSSLab/maia3 repository by the model's actual authors, which is AGPL-3.0-licensed. Their own Hugging Face model cards (MaiaChess/maia3-*) state that the weights follow the repository's license, not an independent one.

These files were originally converted to ONNX and re-uploaded as cemoss17/maia3-onnx under a self-declared MIT tag; this repo previously carried that same MIT tag on the further-modified (int32 Cast node) files here. On review, neither intermediate re-upload documents a relicensing grant from CSSLab, so this repo's license tag has been corrected to AGPL-3.0 to match the authoritative upstream source, pending any clarification from CSSLab.

If you're using these weights: treat AGPL-3.0 obligations as applying, including for network/SaaS use (AGPL's distinguishing term vs. GPL).

Source

Format-converted from cemoss17/maia3-onnx, itself derived from the official CSSLab/maia3 release. Modified here with additional ONNX Cast nodes for int32 input compatibility.

Used by CrispChess, which downloads these weights at runtime (never bundles them) and is working to isolate inference into a separate process/context, matching the treatment CrispChess already gives its other AGPL/GPL-licensed engine options.

Provenance and EU AI Act Art. 53 note

• Upstream model: cemoss17/maia3-onnx — published by cemoss17.
• Upstream licence: agpl-3.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
• What was done here: format conversion and/or quantisation only (ONNX). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
• Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
• Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.