ChessLFM: a 230M-param model hits 2000 Elo and lands on Lichess homepage
maximelabonne · x · 2026-09-22
- Maxime Labonne's ChessLFM, trained from LFM2.5-230M, was republished on Lichess (where he originally collected seed data) and hit the homepage.
- It reaches 2004 Elo (95% CI 1972–2039) on a Stockfish-anchored ladder and runs locally in-browser via WebGPU, beating Fable 5.1 in a demo game.
- Key design: one move = one token, abandoning raw PGN text (chess-GPT reached only 1300–1500 Elo with illegal-move issues), borrowing DeepMind ChessBench's explicit board encoding + Stockfish action-value prediction, plus teacher distillation, light RL, and lots of search.
Related event: 230M-Parameter Chess Model Hits 2000 Elo, Featured on Lichess Homepage(2 posts)→
More from Research
- O'Reilly builds a working data vocabulary for the semantic era, from warehouses to ontologies — rseroter · 2026-09-22
- Glance reads structured visual answers from a frozen 4B VLM, cutting GPU cost up to 85% — multiply_matrix · 2026-09-22
- SkillLift cuts agent skill-evolution token cost 40-70% by ranking, not rollouts — dair_ai · 2026-09-22
- FAIR & NYU scaling law paper pushes fits down to 4M-param models — giffmana · 2026-09-22
- Salesforce: Fine-tuning a weak model to copy Gemini drops success 15%; correcting its own failures works — rohanpaul_ai · 2026-09-22
- Jevons effect sparks encoder model revival: 'BERT revenge' over GPT — JFPuget · 2026-09-22