How a 230M-param model became a 2000 Elo chess engine, playable in your browser
maximelabonne · x · 2026-09-16
Maxime Labonne published a new article showing how to turn a tiny 230M-parameter language model into a roughly 2000 Elo chess engine.
Key ingredients:
- SFT distillation from Stockfish-generated games to teach sensible moves
- A bit of reinforcement learning on top
- Heavy reliance on test-time search to compensate for the small model's limits
A demo (LFM2.5-230M-Chess) runs directly in the browser via Hugging Face; the article link lives in the model card.
More from Research
- HarnessVLN: training-free embodied navigation agent sets SOTA on four benchmarks — Yang Chen · 2026-09-16
- TROT: Tsallis-Regularized Optimal Transport Unifies Wasserstein and KL Divergences — FrnkNlsn · 2026-09-16
- Professor estimates viral post-training algorithms work out of the box only ~5% of the time — Kangwook_Lee · 2026-09-16
- AlpaSim Challenge Borrows LLM Multi-Domain Benchmarking, Uses Item Response Theory for Autonomous Driving Evals — abursuc · 2026-09-16
- CoLLAs 2026 Keynote: Continual Model Merging via Subspace Modeling and Low-Rank Experts — apsarathchandar · 2026-09-16
- Interpretability researcher lists top open problems in decoding model activations — wesg52 · 2026-09-16