LFM2.5-230M-Chess
A chess engine with a language-model interface, by maxime-labonne (Liquid AI) on Liquid's LFM2.5-230M-Base. Read from the model card; arrived by Telegram 2026-09-16. The card points to a Substack write-up and a browser demo (not read).
The adaptation, which is the point
- Every legal chess move is one token. 2,106 tokens were added to the vocabulary, so the model never spells a move character by character and cannot produce a syntactically broken one.
- A position is a fixed 80-token prompt. No PGN history; the board state is the input.
- Two forward passes per move. The first emits value tokens — a win-probability estimate — and the second selects a move, with sampling restricted to the legal set. Illegal moves are excluded by masking, not learned away.
- Trained by Stockfish distillation: the targets are the engine's evaluations and choices.
- Licence
lfm1.0(Liquid's own); an ONNX export is provided.
Strength, first-party
About 1500 Elo as a raw one-pass policy; about 2004 Elo with a depth-3 search wrapped around it. Both figures are the author's, with no stated opponent pool or game count on the card.
Why it is here
This spoke's adaptation corner is mostly PEFT and RL post-training on language tasks. This is a cleaner specimen of a different move: change the vocabulary to fit the task, then distil a teacher that already solves it. The 230M base contributes a sequence model and nothing chess-specific; the 2,106-token move vocabulary and the value-then-policy pass are what make it an engine. It is also a small, checkable case of the distillation claim that a tiny student can carry a large teacher's policy — checkable because Elo is measured by playing, which anyone with the demo can do. The Elo figures are not yet independently measured.
Runner-up spoke: llm-providers (Liquid AI's model line). Routed here because the substance is the fine-tuning recipe, not the provider.
Tier T3 — first-party model card; Elo figures unsourced beyond the author.