Nous Research tunes Qwen3.8-27B with 53M tokens into Arete-27B

On 2026-11-07, NousResearch announced Arete-27B, an open-weight model created by post-training Alibaba's Qwen3.8-27B to adjust its preferences and alignment on politically sensitive topics. The team said open source gives labs the raw tools to mold intelligence to varied preferences, and that the values a model ships with should not be the only values its users get to choose from. It reported seeing refusals, deflections, and CCP-framed answers on politically sensitive topics in Qwen3.8-27B, and post-trained to alter those biases. Later the same day, Nous Research's account on X described Arete-27B as Qwen3.8-27B post-trained to reinforce Western values, adding that Chinese open models are currently the strongest available but ship with state-aligned behavior in the weights, and that the team built an 803-prompt eval spanning seven value categories, including freedom of speech. According to The Decoder's report that day, the fine-tuning relied on 53 million tokens of supervised fine-tuning, and the resulting model's pro-Chinese output rate fell from 22 percent to less than one percent while general capabilities stayed intact in tests. The team writes that this political bias is applied late and applied thin, and can be stripped away by fine-tuning. The record so far covers the launch and these reported measurements; no follow-up release or independent verification is documented in the available reports.