LittlePlyModel + harness

Historical inference record. Browser play has been retired; play LittlePly on Lichess. Validation results below describe the former implementation, not an available website feature.

Historical checkpoint and artifacts; CM-13 is historical; CM-14 is historical; CM-17 is an evaluated checkpoint.

Historical CM-9 browser release. Historical CM-10 browser release.

CM-9 / artifacts and boundaries

Inspect the exact browser release.

The immutable FP32 release is available with source, hashes and validation evidence.

CM-9 checkpoint · Evaluation and limitations · Retired browser artifact · Reproduction instructions

Downloads

Model: 344,416,641 bytes; runtime: 27,190,919 bytes. The browser verifies their assembled hashes before inference. Desktop WebGPU is required; there is no mobile or CPU fallback. Only one NVIDIA Lovelace adapter has been validated. Total GPU memory use is not measured.

Validation and limits

Seven fixtures preserve every legal-move ranking and selected move versus original PyTorch FP32. Maximum browser raw-logit error was 0.00004071. A 111-ply self-play game ended in checkmate and independently replayed legally in Python. Interactive play and reset passed. These are correctness checks, not a playing-strength benchmark.

Full curriculum payloads and resumable optimizer checkpoints are not bundled. This release distributes browser inference weights, not a turnkey training replay. Its training checkpoint identity is recorded separately from the inference and ONNX export hashes. Historical CM-7 artifacts remain available.

Original browser code is Apache-2.0; bundled chess.js and ONNX Runtime notices are preserved alongside the assets. No optimizer state, private game journals or credentials are included.