Nous Research turned Alibaba's Qwen3.8-27B into Arete-27B with supervised fine-tuning on about 53 million tokens over two epochs. The share of answers repeating Beijing's official narrative fell from 22 percent to 0.9 percent, while factual accuracy rose from 2.32 to 2.98 out of 5. For companies evaluating open-weight models, the case suggests political alignment can be removable without rebuilding the base model.
How Nous tested and retrained Qwen3.8-27B
Nous describes Chinese open-weight models as technically strong but carrying state-mandated behavior in their weights. With Qwen3.8-27B, the team observed refusals, evasive answers, and responses aligned with the Chinese Communist Party. The resulting Arete-27B is positioned as carrying Western values instead of Beijing's party line, and its weights, test dataset, and training data are published under Apache 2.0 on Hugging Face.
The project used a custom test of 803 prompts across seven categories, including free speech, state ideology, education, privacy and surveillance, democracy, and environment. Each prompt was tied to two to four facts drawn from a dated source article such as laws, news reports, or research papers. Grok 4.7 from xAI acted as judge, checking whether a response mentioned, omitted, or contradicted those facts without knowing which model produced the answer.
A response counted as propaganda when it reproduced Beijing's official narrative instead of the facts, endorsed it, or presented it as an equally valid perspective. Unmodified Qwen3.8-27B scored 2.32 on factual accuracy, against below one percent propaganda for Mistral Large 4 and Gemma 4 31B, and 15 percent for GLM-5.3 plus 16 percent for DeepSeek V4.1 Flash. Nous selected Inkling from Thinking Machines Lab as teacher because an open model avoided terms that ban distillation, while Nvidia Nemotron 3 Ultra was rejected after irregularities on China-related topics.
What this means for enterprise model selection
For businesses, the practical effect is a wider usable pool of open-weight models. A technically capable model that was previously unsuitable for customer-facing or compliance-sensitive work can become deployable after targeted alignment, expanding vendor choice beyond Western-origin systems. Small firms gain access to Apache 2.0 weights they can download and audit, while larger firms can repeat the same fine-tuning pattern on their own data and evaluation sets.
The limits concern measurement and transfer of the method. The judge was Grok 4.7, itself politically right-leaning, and the reasoning-mode score of 3.07 rests on only two of five runs so far. Training and test overlap was removed, but buyers still need to verify behavior in their languages, response formats, and sensitive topics, since other studies show format and language change results. The news alone does not prove that every Chinese model has an equally thin alignment layer.
The marker to watch is independent replication of the two reported numbers: propaganda below one percent with Hermes Bench near 65.6 percent and Terminal-Bench 2.1 near 70 percent. Nous raised 90 million dollars in October at a 1.5 billion valuation and is moving into enterprise sales with Hermes for Businesses. If enterprise pilots confirm the same stability, thin-layer removal will become a standard step in open-model due diligence.
