Arete-27B: Nous retunes an open model's point of view.

What happened?

Nous Research has released Arete-27B, an open-weight model that, on Nous's own evaluation, states more of the expected facts and far less often repeats the Chinese state's official framing than the Qwen model it was built from.

How it was built

Arete-27B is not a new architecture. It is Alibaba's Qwen3.8-27B, adjusted with supervised fine-tuning: the model learns from example answers written by a teacher model. Here the teacher was Inkling, an open model Nous picked after its own testing.

The run was small next to the original pretraining: about 52.8 million tokens, two passes over the data and 1,202 training steps. The data was text only, so the vision encoder stayed frozen. There was no reinforcement learning, no preference optimisation and no change to the architecture.

How Nous measured it

Nous built an evaluation of 803 English prompts across seven categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment. Each prompt comes with two to four required facts from a dated source article, plus the official language the CCP uses on that topic.

An AI judge, Grok 4.7, grades every answer against a fixed rubric. It records which required facts an answer states, misses or contradicts, and it has to quote the exact phrase whenever it flags an answer for adopting the official line. Nous refined the rubric on a 25-prompt pilot that it graded by hand.

So the two headline numbers measure something specific: recall of targeted facts, and whether an answer takes on one particular framing. They are not a general score for truthfulness.

Why it matters

The practical lesson is that open weights can be adjusted. A fairly small, targeted post-training run changed the behaviour Nous measured in this checkpoint, while the general skills it tested held steady. For teams that build on open models, it is a useful reminder that a model's default answers on sensitive topics are not necessarily fixed.

Just as useful, the evaluation is public. Anyone can run the same prompts and judge setup on Arete-27B, on the original Qwen model or on any other model, and see whether the results hold up with their own settings.

What to keep in mind

These are Nous's numbers, scored by Nous's chosen judge, on prompts Nous designed. "Western values" is Nous's stated target, not a definition of neutrality. A better score on this evaluation is not a guarantee of balanced or safe behaviour in general, and it does not prove that every value in a model sits close to the surface.

Within those limits, Arete-27B is a clear, well-documented example of what post-training on open weights can do, with the tools published so others can check the work.