How to use from
Docker Model Runner
docker model run hf.co/jromanghf/rookstudio-commentary:Q4_K_M
Quick Links

RookStudio commentary model

The language model of RookStudio, a chess game analysis program: it writes short comments on the moves of a game, in English or French, from the facts the program computes with the engine. RookStudio holds it to a grammar and a checker, and falls back to its own templates whenever a comment does not hold to the facts.

  • Base model: Qwen3.5-4B, fine-tuned and merged
  • Files: GGUF Q4_K_M, 2.7 GB each; they run on an 8 GB graphics card or on the processor
File For
commentary-v11-Q4_K_M.gguf RookStudio 0.2.0 and later
commentary-Q4_K_M.gguf RookStudio 0.1.5 and earlier

Each version of the program downloads the model written for it: a new model gets a new file name, and the earlier files stay.

The v11 model

Fine-tuned on 21,700 comments a larger model wrote from the program's facts, prompted exactly as the program prompts it, and kept only when the checker accepted them. It says why a move fails and what the better move does about it, and words the notions of chess books when the facts carry them: a defender drawn away, a pawn majority, the Carlsbad structure, the opposition, the square of the pawn.

On 120 games it was not trained on (2,296 comments to write, English and French), RookStudio's checker accepts 99.7 % of its comments, 98.6 % at the first attempt.

Use

RookStudio downloads it with Ctrl+K โ†’ โ€œDownload the recommended language modelโ€. By hand, put the file in %LOCALAPPDATA%\rookstudio\models\ (Windows) or ~/.local/share/rookstudio/models/ (Linux).

Downloads last month
115
GGUF
Model size
4B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for jromanghf/rookstudio-commentary

Finetuned
Qwen/Qwen3.5-4B
Quantized
(692)
this model