2 days ago

LoGos-7B (Q8_0 GGUF) Local 8-bit GGUF conversion of YichuanMa/LoGos-7B

ollama run ikaankeskin/logos-7b

Details

2 days ago

e122e52590b2 · 8.1GB ·

qwen2
·
7.62B
·
Q8_0
{ "num_ctx": 4096, "temperature": 0.7 }

Readme

LoGos-7B (Q8_0 GGUF)

Local 8-bit GGUF conversion of YichuanMa/LoGos-7B for Ollama and llama.cpp. Used by Kaibitzer.

This is not a new training run. Weights come from the original model; they are stored as Q8_0 so they fit on a consumer GPU. Credit, paper, and Apache-2.0 license follow the original authors.

Hugging Face file: ikaankeskin/logos-7b-gguf

What is LoGos?

LoGos-7B is a 7B language model for Go (weiqi) reasoning: it reads a board position, thinks in a long chain-of-thought, and proposes the next move (color, coordinate, and a win-rate estimate).

It is built on Qwen2.5-7B, then mixed cold-start training plus GRPO so professional Go knowledge and long-CoT reasoning transfer onto actual games. It is a tutor / analysis model, not a replacement for a full-strength engine like KataGo.

Paper: Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go (Ma et al., 2026).

Original weights are BF16 safetensors (~14 GB). This Ollama model is the same weights in Q8_0 GGUF (~8.1 GB).

Quantization

This library model is Q8_0 (8-bit). That is a light quant: small quality drop vs the original BF16, much closer to the original than a 4-bit (Q4) file.

It is not the original Hugging Face safetensors, and it is not Q4. A Q4 built by requantizing this Q8 file would lose more quality.

Verified in Kaibitzer via Ollama on an RTX 3080 Laptop (~8 GB VRAM, ~2.3 tok/s on long CoT). Generation is slow at Q8; that is expected.

Run it

ollama pull ikaankeskin/logos-7b

Kaibitzer defaults: LOGOS_URL=http://127.0.0.1:11434, LOGOS_MODEL=logos-7b. For the Chrome/web client, set OLLAMA_ORIGINS=*.

Same file from Hugging Face:

ollama pull hf.co/ikaankeskin/logos-7b-gguf:logos-7b-q8_0.gguf ollama cp hf.co/ikaankeskin/logos-7b-gguf:logos-7b-q8_0.gguf logos-7b

Prompt

LoGos expects its Chinese training template: move record (1.X-Q16), board matrix (1 black, -1 white, 0 empty), then boxed 下一步位置. Kaibitzer builds that prompt for you. See the original model card for the full template.

Kaibitzer uses context 4096 and num_predict 512.

License

Apache-2.0, same as YichuanMa/LoGos-7B.

@misc{ma2026mixingexpertknowledgebring, title={Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go}, author={Yichuan Ma and Linyang Li and Yongkang Chen and Peiji Li and Jiasheng Ye and Qipeng Guo and Dahua Lin and Kai Chen}, year={2026}, eprint={2601.16447}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2601.16447}, }