Nous Research's Arete-27B is an Apache 2.0 fine-tune of Qwen3.8-27B that cuts state-propaganda answers from 22% to 0.9% on its own eval. What changed, the backlash, and who should try it.
On November 6, 2026, Nous Research announced Arete-27B on X ↗, its fine-tune of Alibaba's Qwen3.8-27B that it says is "post-trained to reinforce Western values." The blog post ↗, the model weights on Hugging Face ↗ and the evaluation set went live together. By November 8 the announcement had more than 14,000 likes, 2 million views and 1,600 replies, making it one of the most argued-about open-weights releases of the week. The Decoder ↗ and MarkTechPost ↗ covered it on November 7.
Nous's argument is simple. Chinese open models such as Qwen are among the strongest you can download, but they carry state-aligned behavior in the weights. Nous observed "refusals, deflections, and CCP-framed answers on politically sensitive topics" in Qwen3.8-27B and set out to measure and remove them.
Per the model card ↗, it's a light fine-tune:
Supervised fine-tuning only. There's no RL, no preference optimization and no change to the architecture.
About 52.8 million tokens at 64k context, over two epochs and 1,202 steps.
The teacher was Inkling from Thinking Machines Lab. Nous says using an open teacher avoided the terms-of-use problems of distilling a closed model. (Inkling's lab just shipped Inkling 2 Small.)
The training data is text-only. Nous froze Qwen's vision encoder rather than retraining it.
The license is Apache 2.0, the same as the base model, per The Decoder and MarkTechPost.
The base model is strong on its own. Alibaba's Qwen3.8-27B card ↗ lists a 27B dense vision-language model with 262,144 tokens of native context, extendable to 1 million.
Nous built an 803-prompt evaluation across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. Each prompt has two to four required facts taken from a dated source. An LLM judge, Grok 4.7, checks whether an answer states, misses or contradicts each fact, and whether it adopts Beijing's official framing. The judge isn't told which model wrote the answer.
Selected rows from Nous's results table:
| Model | Fact score (0–5) | Propaganda rate |
|---|---|---|
| Mistral Large 4 | 3.65 | 0.3% |
| GLM 5.3 | 3.10 | 15% |
| Arete-27B | 2.98 | 0.9% |
| Gemma 4 31B | 2.58 | 0.5% |
| DeepSeek V4.1 Flash | 2.39 | 16% |
| Qwen3.8-27B (base) | 2.32 | 22% |
Nous notes that the decoding settings weren't identical across rows. Arete and Qwen ran at temperature 0, and the others ran at 0.6. The biggest shift came in freedom of expression, where the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%. Answers scoring zero dropped from 359 to 216 out of 803.
On capability, Nous reports Hermes Bench at 65.60% vs 65.77% for the base model, which it calls noise. It also reports Terminal-Bench-2.1 at 70% vs 67%, a gain Nous explicitly declines to take credit for because the harnesses differed. All of these are Nous's own tests. We haven't seen independent replication yet.
The reaction split hard. Developers who run local agents mostly liked the result:
"If you run local agents on Chinese open weights, this is the optionality open models were supposed to give you."
Pratik Gandhi ran both models 4-bit on a 64GB M5 Pro Mac mini and reported 17.2 tok/s for Arete vs 16.8 for Qwen ↗ on TensorFold. In other words, the fine-tune doesn't slow it down. Community GGUF quantizations appeared within a day.
The loudest pushback targeted the framing and the sources. One widely shared critical quote post, from @hot_labor ↗ (740+ likes), set out to "dig deeper into the training dataset" and objected to the sources Nous treats as factual. Others went after the method: "'Reinforce Western values' is a label, not a measurable," wrote @Hanife17 ↗, asking who wrote the rubric. A more measured take came from @techtactician ↗: "That's a striking result, but I'd want independent tests before concluding the underlying bias is gone." Defenders such as @ThatWolfieGuy ↗ argued that "all they did was remove the bias the CCP requires the labs to put in."
The headline isn't the politics. It's how shallow the layer was. Nous's strongest claim is that about 53M tokens of supervised data moved the behavior it measured without hurting the benchmarks it ran. Whatever you think of the "Western values" label, that tells you a model's values are a thin, editable layer on top of its capabilities, in either direction. Anyone deploying open weights should take that seriously.
Who should try it:
Teams already running Qwen3.8-27B locally for research, journalism, education or support work that touches China-related topics. Arete runs at the same speed, and the eval is public, so you can rerun it.
Researchers studying value alignment. You get open weights, an open eval and a documented recipe. That's rare.
Who should skip it, or wait:
Anyone who needs a neutral assistant. Nous says plainly that it swapped one viewpoint for another one it prefers. If your product serves users across regions, test it on your own prompts first.
Anyone shopping on capability. Arete isn't smarter than Qwen3.8-27B. If politics doesn't come up in your workload, the base model, DeepSeek V4.1 or Gemma 4 will do the same job.
For the wider open-weights shortlist, including Qwen, Muse Spark Open and Mistral Large 4, see our best open-source LLMs of 2026. Other open releases from this week include Kimi K3.1 Pro and Baidu ERNIE-5.5-Flash-Base. If you're wiring Arete into an agent stack, our agent frameworks comparison covers the harness side.
Nous Research: Introducing Arete-27B ↗, November 2026
Hugging Face: NousResearch/Arete-27B model card ↗, checked November 8, 2026
Hugging Face: Qwen/Qwen3.8-27B model card ↗, checked November 8, 2026
The Decoder: State propaganda in some Chinese AI models is only "a thin layer" ↗, November 7, 2026
MarkTechPost: Nous Research releases Arete-27B ↗, November 7, 2026