Browser weights: CM-7, with a three-ply helper (50,000 nodes / one second). Lichess uses the same checkpoint with a separate adaptive seven-ply helper. CM-7 uses the state-input format introduced by CM-6.

Desktop WebGPU / Experimental

Your browser. LittlePly’s next move.

Play against CM-7 weights locally. Your desktop GPU runs the model; no game moves are sent to an inference server.

Experimental desktop WebGPU only (Chrome or Edge with a compatible GPU). Validated on one NVIDIA Lovelace adapter; other GPUs are unverified. No mobile or CPU fallback. FP32 weights: 346.8 MB, plus 27.2 MB runtime and JavaScript. GPU memory is not measured and exceeds the download size. First load may take time; caching is best effort.

Prefer Lichess? Play LittlePly there · Inspect browser source · Model manifest and hashes

Your pieces

Model/runtime files download only after you choose to load.

Checking desktop WebGPU support.

Your game

Choose a piece, then a highlighted destination. Keyboard: Tab to a square, Enter to select it. Changing sides resets the game.

Promote to

What is running?

The selected 86.24-million-parameter CM-7 checkpoint, current board/state plus optional last 16 genuine plies, and three-token move codec. Legal candidates are ranked by model likelihood. A separately bounded three-ply mate helper can override the model; overrides are labeled and recorded in exported PGNs. The model prompt stays fixed at 120 slots, with up to two candidate tokens. Full game history stays in the controller for legality, repetition and PGN export.

Games stay in this tab unless you export them; they are lost on refresh. Static assets are requested from this site after consent for the model download. Site visit/performance analytics remain enabled, as disclosed here; this is not a claim of zero network traffic. The game is unrated and separate from your Lichess account.

Browser portability and memory limits are experimental. If loading fails, reset/cancel, then retry or use the Lichess link. Reset during download or inference terminates the worker and requires reloading the model. No continuous-availability or universal-speed claim.

ONNX Runtime license · chess.js license · Original code Apache-2.0