Nous Research releases Arete-27B, Qwen3.8-27B retuned to drop Chinese propaganda

Nous Research, a US startup best known for its open-source agent Hermes, has released Arete-27B, an open-weight model built by fine-tuning Alibaba's Qwen3.8-27B. The model promises "Western values" instead of Beijing's party line. According to Nous, Chinese open-weight models rank among the strongest available but carry state-mandated behavior baked into their weights. With Qwen3.8-27B, the team observed refusals, evasive answers, and responses aligned with the Chinese Communist Party.

To measure this, Nous built its own test with 803 prompts across seven categories, including free speech, state ideology, education, privacy and surveillance, democracy, and the environment. For each prompt the team defined two to four facts drawn from a dated source article: laws, news reports, or research papers. The judge model is Grok 4.7 from Elon Musk's xAI, described by the source as politically right-leaning, which checks whether a response mentions, omits, or contradicts those facts. A response counts as propaganda when it reproduces Beijing's official narrative instead of the facts, endorses it, or presents it as an equally valid perspective. The judge must cite the specific wording and does not know which model wrote the answer.

On this test the unmodified Qwen3.8-27B scored 2.32 out of 5 on factual accuracy, and 22 percent of its answers contained propaganda. Western models such as Mistral Large 4 and Gemma 4 31B came in below one percent, while GLM-5.3 and DeepSeek V4.1 Flash landed at 15 and 16 percent.

Nous chose Inkling from Thinking Machines Lab as the teacher model, because an open model would not break terms of service that ban distillation. Nvidia's Nemotron 3 Ultra was also on the shortlist, but Nous says there were signs of irregularities on China-related topics. The team suspects the bias carried over from training on Chinese open-weight models, and it reports similar contamination in its own projects. The source adds that Thinking Machines itself used models including the Chinese Kimi K2.5 for Inkling's synthetic training data.

Training was supervised fine-tuning only, on about 53 million tokens over two epochs, with no reinforcement learning or preference optimization. The team says it removed overlaps between training and test prompts beforehand. Afterwards Arete-27B scores 2.98 on factual accuracy and the share of propagandistic answers drops to 0.9 percent. The biggest shift was in the free speech category, where the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60 to one percent. With reasoning mode on, Arete reaches 3.07, though that figure rests on only two of five runs so far.

Nous says general capabilities do not suffer. On its internal Hermes Bench, Arete-27B scores 65.60 percent against 65.77 percent for the base model. On Terminal-Bench 2.1 the score rises from 67 to 70 percent, but Nous does not call that a real gain because the test setups differed. The team concludes "that the bias layer in a Chinese-aligned model is shallow," and the company writes that it is applied late and thin, and that fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath.

The source frames Arete as a change of direction. Hermes 3, in 2024, was deliberately neutral and designed to adapt to whatever worldview was set in the system prompt. Bias in Chinese models is well documented, it notes. In 2025 NewsGuard found that five leading models, including Qwen, repeated or failed to correct pro-Chinese disinformation in 60 percent of cases. CrowdStrike found that DeepSeek produces less secure code when queries reference Falun Gong, Tibet, or Taiwan. An Aleph Alpha study with 967 politically sensitive questions found that among models from Alibaba, DeepSeek, and Moonshot AI only 17 to 41 percent of answers were balanced. A recent ValsAI study shows that Chinese models differ: GLM-5.3 Flash explicitly confirmed human rights violations, while Kimi K3 denied the existence of internment camps. Language and response format had a clear effect, and even Western models occasionally adopted official Beijing positions. There, too, targeted fine-tuning proved helpful.

Nous Research raised $90 million in a Series B round in October at a $1.5 billion valuation, led by Robot Ventures with Nvidia and Samsung among the participants. It is using the money to move into enterprise sales with "Hermes for Businesses." The weights, test dataset, and training data for Arete-27B are available under an Apache 2.0 license on Hugging Face.

Key facts

• Nous Research fine-tuned Alibaba's Qwen3.8-27B with supervised fine-tuning on about 53 million tokens over two epochs, with no reinforcement learning or preference optimization, producing the open-weight Arete-27B.
• On Nous's own 803-prompt, seven-category test, the propaganda rate fell from 22 percent to 0.9 percent and factual accuracy rose from 2.32 to 2.98 (out of 5 for the base model); in the free speech category the propaganda rate fell from 60 to one percent.
• Nous says general capabilities held: 65.60 percent on its internal Hermes Bench versus 65.77 percent for the base model.
• Nous concludes the bias layer in a Chinese-aligned model is shallow; Inkling from Thinking Machines Lab served as the teacher model.
• Weights, test dataset, and training data are published under an Apache 2.0 license on Hugging Face.

Why it matters

Nous argues that the political slant in strong Chinese open-weight models is a thin, late-applied layer rather than something woven through the model, and that about 53 million tokens of supervised fine-tuning can remove it. If that holds, anyone building on these models can correct the slant cheaply instead of avoiding the models. The source frames Arete as a shift for Nous, whose Hermes 3 in 2024 was deliberately neutral and adapted to whatever worldview the system prompt set. It also sits against a body of earlier findings: NewsGuard in 2025 (five leading models, including Qwen, repeated or failed to correct pro-Chinese disinformation in 60 percent of cases), CrowdStrike on less secure DeepSeek code for Falun Gong, Tibet or Taiwan queries, and Aleph Alpha, where only 17 to 41 percent of answers from Alibaba, DeepSeek and Moonshot AI models were balanced.

Who it affects

Developers and companies that use or would like to use Chinese open-weight models such as Qwen and worry about state-aligned answers. Researchers studying model bias also get the test dataset and training data. Nous itself is affected commercially: it raised $90 million in a Series B in October at a $1.5 billion valuation, led by Robot Ventures with Nvidia and Samsung among the participants, and is moving into enterprise sales with "Hermes for Businesses."

How to use it

The weights, test dataset, and training data for Arete-27B are available under an Apache 2.0 license on Hugging Face. Anyone can run the model, rerun the 803-prompt test, or reuse the training data. Per the source, a reasoning mode exists and gives a higher factual score (3.07), but that number rests on two of five runs so far.

How solid is it

The headline numbers come from Nous's own test and its own benchmarks, and the source reports no independent replication. The judge model is Grok 4.7, which the source describes as politically right-leaning; the source notes this ironically and gives no analysis of how it affects results. The team says it removed overlaps between training and test prompts beforehand. The 3.07 reasoning score is based on only two of five runs. The Terminal-Bench 2.1 rise from 67 to 70 percent is not claimed by Nous as a real gain, because the test setups differed. The claim that general capabilities are intact is Nous's, based on Hermes Bench (65.60 versus 65.77 percent). The source does not say how the 53 million tokens were generated beyond the use of Inkling as teacher model.

Risks and caveats

The 22 percent to 0.9 percent drop is measured on one test whose definition of propaganda (reproducing Beijing's official narrative, endorsing it, or presenting it as equally valid) was written by Nous, and the model is marketed as promising "Western values," so it replaces one alignment with another. Chinese models also differ: ValsAI found GLM-5.3 Flash explicitly confirmed human rights violations while Kimi K3 denied the existence of internment camps, and language and response format affected results; even Western models sometimes adopted official Beijing positions. Nous also reports signs of China-related irregularities in Nvidia's Nemotron 3 Ultra and says similar contamination appears in its own projects, so the teacher model choice matters. The source does not say whether other Qwen versions or other Chinese models were treated the same way.

“It is applied late, it is applied thin, and fifty million tokens of carefully chosen supervision removes it without disturbing anything underneath”

— Nous Research, on the bias layer in Chinese-aligned models