Jeffrey Cheng introduced QUEEN on Oct. 5 as a chess model that chooses moves and explains them in natural language. His announcement describes a 4-billion-parameter model approaching a typical grandmaster’s playing strength.
What the benchmark measures
The authors’ preprint combines a Leela Chess Zero encoder with SmolLM3-3B, then trains on explanations of candidate moves and their continuations. Seven rounds of search distillation raised its estimated rating from 1,782 to 2,697.
That estimate comes from 32 games against eight engines, calibrated to the Lichess Elo scale. The paper compares it with a median grandmaster’s Lichess blitz rating of 2,730. It is not a formal grandmaster title or FIDE rating.
Explanations remain fallible. A language-model judge scored QUEEN’s conceptual coherence below the tested frontier models, and the authors warn that mistaken chess motifs can propagate through training.
Code and documentation
The public repository contains training, data-generation and evaluation code, and links to model downloads. Its README promises fuller documentation and a cleaned code release. It also cautions that the combined GPU inference setup is still being validated.
Comments