Nous Research has released Arete-27B, an open-weight SI model fine-tuned to remove state-mandated biases found in Chinese large language models. The US startup reports that targeted supervision can strip away political alignment layers without degrading general capabilities, marking a shift in how developers approach the neutrality of open-source SI systems.

What Happened

Nous Research fine-tuned Alibaba’s Qwen3.8-27B using approximately 53 million tokens of data to create Arete-27B. According to the company, the original model exhibited refusals and evasive answers aligned with the Chinese Communist Party on sensitive topics. The fine-tuning process involved supervised learning over two epochs, avoiding reinforcement learning or preference optimization. Nous Research selected Thinking Machines Lab’s Inkling as the teacher model for the training data, noting that other candidates like Nvidia’s Nemotron 3 Ultra showed irregularities on China-related topics, possibly due to contamination from Chinese open-weight models during their own training.

To evaluate the results, Nous developed a test suite of 803 prompts covering seven categories, including free speech, state ideology, privacy, and democracy. The company used Elon Musk’s xAI Grok 4.7 as a judge model to assess whether responses mentioned, omitted, or contradicted specific facts drawn from dated sources. The evaluation defined propaganda as reproducing Beijing’s official narrative instead of factual reporting, or presenting it as an equally valid perspective. Unmodified Qwen3.8-27B showed a propaganda rate of 22 percent and a factual accuracy score of 2.32 out of 5. In contrast, Arete-27B achieved a factual accuracy of 2.98 and reduced the propaganda rate to 0.9 percent.

Why It Matters

The findings suggest that political alignment in some Chinese SI models is a shallow layer applied late in training, rather than a deep structural bias. Nous Research states that the bias layer is "applied thin," and that fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath. On Nous’s internal Hermes Bench, Arete-27B scored 65.60 percent, comparable to the base model’s 65.77 percent. On Terminal-Bench 2.1, the score rose from 67 to 70 percent, though the company attributes this to differences in test setups rather than a genuine capability gain.

This development is significant for the open-source SI ecosystem, where Chinese models like Qwen, GLM, and DeepSeek are among the strongest available. Previous studies, including those by NewsGuard and Aleph Alpha, have documented pro-Chinese disinformation and unbalanced answers in these models. For instance, a 2025 NewsGuard report found that five leading models repeated or failed to correct disinformation in 60 percent of cases. CrowdStrike also noted that DeepSeek produced less secure code when queries referenced sensitive topics like Falun Gong or Tibet. By demonstrating that these biases can be technically mitigated, Nous offers a pathway for Western developers to leverage the performance of Chinese open weights while enforcing different value alignments.

The Bottom Line

Nous Research has released the weights, test dataset, and training data for Arete-27B under an Apache 2.0 license on Hugging Face. The company, which raised $90 million in a Series B round in October at a $1.5 billion valuation, is using this release to expand its enterprise offerings. While the propaganda rate drops to levels comparable to Western models like Mistral Large 4 and Gemma 4 31B, the study highlights the ongoing need for rigorous evaluation of political bias in open-source SI systems.