Maia-3 5M β€” LiteRT (.tflite) conversion

A LiteRT (TensorFlow Lite) conversion of the released Maia-3 5M checkpoint by the University of Toronto CSSLab (code). It predicts which move a human of a given rating would play. It adds no training data and does not change the learned weights.

The web build of compose-multiplatform-chess downloads it from this repo. Its Android and iOS builds bundle the same file.

Licence

GNU Affero General Public License v3.0, the licence of Maia-3's code and weights.

File

File maia3-5m.tflite
Size 20,813,796 bytes
SHA-256 f0acad2d11db917c10ecc4418a5ef5600c00e21dcac1f54b77008acfeeed998e
Source checkpoint UofTCSSLab/Maia3-5M @ b6559de2398d7140b985f28fd2c19fb5e47ddabe, maia3-5m.pt, SHA-256 ba14208b2992d85502f5fb501934abf6aaaeb355e9f3fdf90e326911f562524f
Maia-3 source CSSLab/maia3 @ 1e13597c42d4858b7cfd7cfdae01e297263364b2
Converter litert-torch 0.10.0 on torch 2.14.1, with jax/jaxlib 0.11.2

Signature serving_default:

  • Inputs:
    • args_0: board tokens, float32 [1, 64, 97]. That is eight 12-plane one-hot boards (oldest first, each mirrored and colour-swapped when Black is to move) plus a zero clock column.
    • args_1: self Elo, int32 [1].
    • args_2: opponent Elo, int32 [1].
  • Output: output_0, move logits [1, 4352]:
    • indices 0–4095 are from * 64 + to;
    • indices 4096–4351 are promotions at 4096 + (fromFile * 8 + toFile) * 4 + {q, r, b, n}.

The only graph change is that torch.nn.RMSNorm is written out as x * rsqrt(mean(xΒ²) + eps) * weight, with torch's default eps. Every legal-move logit matches the unmodified PyTorch model within 9.5e-06 on the reference positions.

Reproduction

The pinned environment and the export script are in the app repository: tools/maia3/export_litert.py and tools/maia3/requirements-litert.txt. The full record is in docs/maia3-model-provenance.md. The export is deterministic: re-running it reproduces this file byte for byte.

Downloads last month
11
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for ber4444/maia3-5m-litert

Finetuned
(2)
this model