THE ESSENTIALS IN 30 SECONDS

  • Arete-27B is a supervised fine-tune of Qwen3.8-27B, trained on about 52.8 million tokens with no reinforcement learning or preference optimization.
  • On Nous's 803-prompt evaluation, judged by Grok 4.7, answers containing propaganda fell from 22% to 0.9% and the average fact score rose from 2.32 to 2.98 out of 5.
  • The results are self-reported, the benchmark and judge setup are Nous's own, and the company says it has not fixed every related behaviour.

Nous Research published Arete-27B on 6 November, a 27-billion-parameter model built on Alibaba's Qwen3.8-27B. The lab says the base model tends to refuse, deflect or echo Chinese government framing on politically sensitive questions, and describes its goal as reinforcing "western values" in a model trained in China. The weights are on Hugging Face under the Apache 2.0 licence, alongside the training pairs and the evaluation set.

How it was trained

According to the model card, Arete-27B received supervised fine-tuning only: no reinforcement learning, no preference optimization and no change to the architecture. Nous generated about 20,000 prompts across seven themes, including freedom of expression, state ideology, democracy, and data privacy and surveillance, then removed duplicates and any overlap with the test set. Responses came from a teacher model called Inkling, which The Decoder identifies as a large open model from Thinking Machines Lab, mixed with answers from Qwen itself. Training ran for two epochs over roughly 52.8 million tokens. The vision encoder was frozen, as the data was text only.

What Nous claims

Nous built an 803-prompt test with facts drawn from dated sources and used xAI's Grok 4.7 as the grader. On that test, the share of answers containing propaganda fell from 22% to 0.9%, and the average fact score went from 2.32 to 2.98 out of 5. The biggest shift was on freedom of expression, where propaganda dropped from 60% to about 1%. On Nous's internal Hermes Bench, the model scored 65.60% against 65.77% for the base, which the lab presents as no loss of general ability. A three-point rise on Terminal-Bench 2.1 is not counted as a real gain, because setups differed.

Caveats

All figures are Nous's own. The benchmark, the LLM judge and the rubric were designed in-house, and the judge was calibrated on a hand-graded pilot of only 25 prompts. Decoding settings vary between rows of the comparison table, and a result with reasoning turned on rests on two of five runs. The lab says it addressed what it considers a root cause rather than every biased behaviour, and it openly frames the choice of target values as a normative one. The Decoder's coverage cites no independent check of the results so far.