Versions
Choose v2 for MTP speculative decoding and FP8. Choose v3 for the lowest censorship rate.
Comparison with Arete-27B (Nous Research)
Arete-27B is a fine-tune of the same base model that also targets political bias. Measured on the BF16 release. We ran it through the same pipeline: identical prompts, decoding (non-thinking, greedy) and judge for all three models. Percentages are of all 115 sensitive questions; the values row uses 10 follow-up questions that ask the model for its own judgment on contested policies.

Table with columns: Metric, Qwen3.8-27B, Arete-27B, Uncensored-v2| Metric | Qwen3.8-27B | Arete-27B | Uncensored-v2 |
|---|
| Censored answers (lower is better) | 83% | 50% | 3.5% |
| Refuses or evades (lower is better) | 50% | 23% | 1.7% |
| Rewrites in CCP talking points (lower is better) | 32% | 28% | 1.7% |
| Sides with CCP values (lower is better) | 100% | 30% | 10% |
| Correct subject, answered openly (higher is better) | 12% | 38% | 77% |
| MTP head for speculative decoding | Yes | No | Yes |
The values component of Uncensored-v2 was fitted on Pan & Xu prompts, so the values row is in-sample for Uncensored-v2 and out-of-sample for Arete-27B.
Evaluation
The settings are identical to those described on the BF16 model card.
1. Political censorship (Pan & Xu, 2026)
113 of the 115 politically sensitive questions (No. 6 and No. 109 excluded as knowledge gaps of the base model; see the BF16 card); non-thinking mode, greedy decoding, 4,096 maximum output tokens; each response labeled as censored or uncensored by an LLM judge.
Table with columns: Model, Censored responses (of 113; Q6 and Q109 excluded), Censorship rate| Model | Censored responses (of 113; Q6 and Q109 excluded) | Censorship rate |
|---|
| Qwen3.8-27B | 98 | 86.7% |
| Qwen3.8-27B-Uncensored-v2-FP8 | 3 | 2.7% |
Uncensored-v2-FP8 removes 96.9% of the censored responses of Qwen3.8-27B (95 of 98).
Including the two excluded questions, censored responses on all 115 sensitive questions are 100 for Qwen3.8-27B and 4 for Qwen3.8-27B-Uncensored-v2-FP8.
2. Refusal of harmful requests
First 100 prompts of the mlabonne/harmful_behaviors test split (AdvBench-derived).
Table with columns: Model, Harmful requests refused (of 100)| Model | Harmful requests refused (of 100) |
|---|
| Qwen3.8-27B | 100 |
| Qwen3.8-27B-Uncensored-v2-FP8 | 97 |
3. General capability
Capability benchmarks were run on the BF16 release:
Table with columns: Benchmark, Reference, Reference score, Uncensored-v2 (BF16), Difference| Benchmark | Reference | Reference score | Uncensored-v2 (BF16) | 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
vllm serve InterpreticAI/Qwen3.8-27B-Uncensored-v2-FP8 --reasoning-parser qwen3
We recommend the sampling parameters published for Qwen3.8-27B: 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.
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 and Qwen/Qwen3.8-27B-FP8. 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
Citation
@misc{interpretic2026qwen3_8_27b_uncensored_v2_fp8, title = {Qwen3.8-27B-Uncensored-v2-FP8}, year = {2026}, howpublished = {\url{https://huggingface.co/InterpreticAI/Qwen3.8-27B-Uncensored-v2-FP8}}}