library_name: transformers
license: apache-2.0
pipeline_tag: image-text-to-text
base_model:
- NousResearch/Arete-27B
- Qwen/Qwen3.8-27B
tags: - qwen3.8
- uncensored
- political-censorship
Qwen3.8-27B-Uncensored-v3
Qwen3.8-27B-Uncensored-v3 is built on NousResearch/Arete-27B, a post-trained Qwen3.8-27B. Arete-27B already removes most propaganda on English-language prompts. On Chinese-language political questions, however, it still censors about half of its answers. Uncensored-v3 removes almost all of that remaining censorship. It is a full BF16 checkpoint with the same architecture and file layout as Arete-27B.
Evaluation
All three models were evaluated under identical settings: same prompts, same hardware and inference engine, non-thinking mode, greedy decoding, and the same LLM judge.
1. Political censorship on Chinese-language prompts (Pan & Xu, 2026)
We use the 145-question Chinese-language political battery of Pan and Xu (2026) and report results on all 115 politically sensitive questions, with a limit of 4,096 output tokens. The judge labels each answer as a refusal or evasion, an answer in CCP talking points, or an uncensored answer. Ten follow-up questions ask the model for its own judgment on contested CCP policies.
| Metric | Qwen3.8-27B | Arete-27B | Uncensored-v3 |
|---|---|---|---|
| Censored answers (lower is better) | 83.5% | 50.4% | 2.6% |
| Refuses or evades (lower is better) | 49.6% | 22.6% | 0.9% |
| Rewrites in CCP talking points (lower is better) | 32.2% | 27.8% | 1.7% |
| Sides with CCP values (lower is better) | 100% | 30% | 20% |
| Correct subject, answered openly (higher is better) | 12.2% | 38.3% | 80.9% |
Uncensored-v3 removes 95% of the censored answers that remain in Arete-27B.
2. Propaganda on English-language prompts (Nous Research Propaganda_Eval)
We use the 803 prompts of NousResearch/Propaganda_Eval with its grading and scoring scripts unchanged, and one grader for all three models.
| Model | Answers containing propaganda (lower is better) |
|---|---|
| Qwen3.8-27B | 24.5% |
| Arete-27B | 1.0% |
| Qwen3.8-27B-Uncensored-v3 | 0.5% |
Usage
The model is used exactly as Arete-27B, including its chat template and its thinking and non-thinking modes. 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.
vllm serve InterpreticAI/Qwen3.8-27B-Uncensored-v3 --reasoning-parser qwen3
As in Arete-27B, the multi-token prediction (MTP) module is not included, so serve the model with standard decoding. If you need MTP, use Qwen3.8-27B-Uncensored-v2, which keeps it. The vision encoder is identical to that of Arete-27B. All evaluations reported here are text-only.
Limitations
- Each model was generated once per benchmark.
- A substantive answer is not necessarily a correct one. Users should verify factual claims independently.
License
Apache 2.0. This model is derived from NousResearch/Arete-27B and Qwen/Qwen3.8-27B, both released under Apache 2.0. The license and attribution of Nous Research and the Qwen team are retained; see the LICENSE file.
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
Nous Research (2026). Arete. https://nousresearch.com/arete
Citation
@misc{interpretic2026qwen3_8_27b_uncensored_v3,
title = {Qwen3.8-27B-Uncensored-v3},
year = {2026},
howpublished = {\url{https://huggingface.co/InterpreticAI/Qwen3.8-27B-Uncensored-v3}}
}