Nous releases Arete-27B, a Qwen3.8-27B tune it says cuts state-line answers to 0.9%

Nous Research released Arete-27B on November 6, a post-trained version of Alibaba's open-weights Qwen3.8-27B. Nous says it trained the model “to reinforce western values on a Chinese-trained model.”

The measurement is Nous's own. Its evaluation set contains 803 prompts spanning seven value categories, graded by an LLM judge, Grok 4.7, against required facts and a fixed rubric without being told which model produced an answer. Nous reports that the share of answers labeled “Any propaganda” — ones adopting the Chinese state's account — fell from 22% for the starting model to 0.9%, while the fact score rose from 2.32 to 2.98 out of 5.

By simple arithmetic that is a roughly 96% cut in flagged answers, and the number of answers scoring zero dropped from 359 to 216 out of 803.

The training recipe was deliberately modest: supervised fine-tuning only, with no reinforcement learning, no preference optimization and no architectural change, over about 52.8 million tokens for two epochs. The teacher was Thinking Machines' open Inkling model, which Nous described as the best open model on its target and said avoided the terms-of-use restrictions attached to distilling from closed systems.

In Nous's comparison table, Inkling scored 0.3%, Mistral Large 4 scored 0.3%, GLM 5.3 scored 15% and DeepSeek V4.1 Flash 16%, the last two with thinking turned off. Nous says the change came without meaningful regression on the general-capability and agentic benchmarks it ran.

The results come with caveats. The same team built the model and the test, and the numbers are self-reported. Decoding settings also differed: Arete-27B and Qwen3.8-27B ran at temperature 0 while the comparison models ran at 0.6 — a gap Nous acknowledges in its own write-up.

No outside group has reproduced the results. Nous says it “addressed a potential root cause” but does not claim to have fixed every behavior, and the report says Alibaba has not commented.

The release matters as a demonstration: without retraining the base model or changing its architecture, supervised fine-tuning alone can visibly shift an open model's value posture. That makes value alignment cheaper for anyone customizing open weights — and sharpens the old question of how much a benchmark is worth when the model and the exam come from the same authors.