ConceptCountry releases qwen-concept-27b, a weight-edited Qwen3.8-27B
Original title: ConceptCountry/qwen-concept-27b — new model trending #30 on Hugging Face
Authors report GPQA Diamond up 4.3 and LiveCodeBench v6 up 13.0, with thinking tokens down 27-54%.
ConceptCountry published qwen-concept-27b, a Qwen3.8-27B variant with nine down-projections rewritten by weight-space programming. The edit is derived from the model's own activations, with no gradient steps or training data. Against the unmodified base, the authors report GPQA Diamond +4.3 and LiveCodeBench v6 +13.0, plus 27-54% fewer thinking tokens. A norm-matched random edit did not improve scores. The release is experimental and Apache-2.0.
Unconfirmed: Benchmark gains are the authors' reported results and have not been independently confirmed.
- Weights ship as bf16, FP8, NVFP4 W4A16, Q8_0 and Q4_K_M.
- Reported scores come from bf16 weights served at Q8_0.
- The 4-bit conversion has not been separately evaluated.
- Each folder includes a SHA256SUMS file.
- License is Apache-2.0, inherited from Qwen3.8-27B.
- Vendors
- Alibaba Cloud
- Organizations
- ConceptCountry
- AI models
- qwen-concept-27bQwen3.8-27B
From the original article
1 min read at huggingface.coQwen3. 8-27B with nine down-projections rewritten by weight-space programming. No gradient steps, no training data: the edit is derived from the model's own The model thinks less and scores higher.
By ConceptCountry · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.