library_name: transformers
license: apache-2.0
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.8-27B
tags:
- qwen3.8
- uncensored
- political-censorship
Qwen3.8-27B-Uncensored-v2
Qwen3.8-27B-Uncensored-v2 is a variant of Qwen3.8-27B with political censorship largely removed. It retains the base model's refusal of harmful requests and remains close to the base model on standard reasoning and instruction-following benchmarks. The release is a full BF16 checkpoint with the same architecture and file layout as Qwen3.8-27B.
Other formats: FP8 · GGUF (Q4_K_M)
Evaluation
1. Political censorship (Pan & Xu, 2026)
We use the 145-question Chinese-language political battery of Pan and Xu (2026) and report results on its 113 politically sensitive questions. Responses are generated in non-thinking mode with greedy decoding and a limit of 4,096 output tokens. Following the censorship audit of Yang et al. (2026), an LLM judge labels each response as censored or uncensored.
| Model | Censored responses (of 113) | Censorship rate |
|---|---|---|
| Qwen3.8-27B | 98 | 86.7% |
| Qwen3.8-27B-Uncensored-v2 | 4 | 3.5% |
Uncensored-v2 removes 95.9% of the censored responses of Qwen3.8-27B (94 of 98).
2. Refusal of harmful requests
We measure refusals on the first 100 prompts of the test split of mlabonne/harmful_behaviors, an AdvBench-derived set of harmful instructions, under the system prompt "You are a helpful assistant." A response counts as a refusal if it contains a standard refusal marker.
| Model | Harmful requests refused (of 100) |
|---|---|
| Qwen3.8-27B | 100 |
| Qwen3.8-27B-Uncensored-v2 | 95 |
3. General capability
Thinking mode, reasoning effort xhigh, 65,536 maximum output tokens, temperature 1.0, top-p 0.95, top-k 20, one sample per question. GPQA Diamond and IFBench are compared with the scores reported on the official Qwen3.8-27B model card. MMLU is not reported on that card, so we compare with Qwen3.8-27B evaluated under identical settings on the same 200 randomly sampled questions.
| Benchmark | Reference | Reference score | Uncensored-v2 | Difference |
|---|---|---|---|---|
| GPQA Diamond | Qwen3.8-27B (official card) | 89.2 | 86.4 | −2.8 |
| IFBench | Qwen3.8-27B (official card) | 79.5 | 79.0 | −0.5 |
| MMLU (200 questions) | Qwen3.8-27B (our evaluation) | 87.5 | 86.5 | −1.0 |
Usage
The model is used exactly as Qwen3.8-27B, including its chat template and its thinking and non-thinking modes. We recommend the sampling parameters published for the base model: for thinking mode, temperature=1.0, top_p=0.95, top_k=20; for non-thinking mode, temperature=0.7, top_p=0.8, top_k=20, presence_penalty=1.5.
vllm serve InterpreticAI/Qwen3.8-27B-Uncensored-v2 --reasoning-parser qwen3
The vision encoder is identical to that of the base model. All evaluations reported here are text-only.
Limitations
- The political evaluation covers Chinese-language questions only.
- A substantive answer is not necessarily a correct one. Users should verify factual claims independently.
License
Apache 2.0, inherited from Qwen/Qwen3.8-27B. The license and attribution of the Qwen team are retained.
References
Pan, J., & Xu, X. (2026). Political censorship in large language models originating from China. PNAS Nexus, 5(2), pgag013. https://doi.org/10.1093/pnasnexus/pgag013
Yang, G., Liu, F., Wang, A., Hosseinmardi, H., & Ghasemian, A. (2026). How China-origin vision-language models move from refusal to reframing in state alignment. arXiv preprint arXiv:2608.11816. https://doi.org/10.48550/arXiv.2608.11816