Nous Research Fine-Tunes Alibaba's Qwen to Strip Out Chinese State Propaganda

The model, its evaluation set and its training data are available on Hugging Face under an Apache 2.0 licence. "We post-trained over Alibaba's Qwen3.8-27B to reinforce western values on a Chinese-trained model," Nous wrote on the model card, adding that "the values a model ships with should not be the only values its users get to choose from."

The work goes to one of the main objections companies raise about Chinese open-weight models, which rank among the strongest freely available systems. Menlo Ventures found this week that Chinese models take 89% of enterprise spending on open-source AI. Nous argues that the political alignment in such models does not run deep. The company concluded that "the bias layer in a Chinese-aligned model is shallow", according to The Decoder, and that roughly fifty million tokens of chosen examples remove it without disturbing the rest of the model.

How the Model Was Retrained

According to the model card, Nous used supervised fine-tuning only, with no reinforcement learning, no preference optimisation and no changes to the architecture. Training ran for two epochs and 1,202 steps on about 52.8 million tokens, drawn from roughly 20,000 text prompts across seven value categories.

The teacher model that generated the target answers was Inkling, from Thinking Machines Lab. The Decoder reported that Nous chose it partly because an open model would not breach terms of service that ban distillation. Nvidia's Nemotron 3 Ultra was also considered, but Nous said it saw irregularities in that model's answers on China-related topics and suspects the bias carried over from training on Chinese open-weight models. Thinking Machines itself used models including Moonshot AI's Kimi K2.5 to produce synthetic training data for Inkling, The Decoder noted.

Measured by Its Own Test

Nous built its own benchmark of 803 prompts covering areas including free speech, state ideology, education, privacy and surveillance, democracy and the environment. Each prompt was paired with two to four facts taken from dated laws, news reports or research papers. xAI's Grok 4.7 acted as judge without knowing which model wrote each answer. It checked whether a response stated, left out or contradicted those facts, and it had to quote the wording behind each verdict. An answer counted as propaganda if it repeated Beijing's official account in place of the facts, endorsed it, or presented it as an equally valid view.

On that test, Arete-27B's factual accuracy score rose from 2.32 to 2.98 out of 5. The largest change came in the free speech category, where the propaganda rate fell from 60% to 1%. For comparison, Nous put Mistral Large 4 and Gemma 4 31B below 1%, and GLM-5.3 and DeepSeek V4.1 Flash at 15% and 16% respectively.

The company says the retraining did not cost general ability. On its internal Hermes Bench, the model scored 65.60% against 65.77% for the base model, a gap the model card calls noise. Its Terminal-Bench 2.1 score rose from 67% to 70%, but Nous does not claim the gain because the two runs used different setups. A higher factual score of 3.07 with reasoning switched on rests on only two of five planned runs so far, according to The Decoder.

All of these results come from a benchmark Nous designed and a single commercial model acting as judge, and none has yet been independently replicated.

The release also changes Nous's position. Its Hermes 3 model in 2024 was built to be neutral and take on whatever worldview a user set in the system prompt. Arete takes an explicit stance.

Nous raised $90 million in a Series B round in October at a $1.5 billion valuation, led by Robot Ventures with Nvidia and Samsung taking part. It is using the money to sell to companies through a product called Hermes for Businesses.