Nous Research released Arete-27B on 7 November, an Apache 2.0-licensed fine-tune of Alibaba's Qwen3.8-27B, retuned to answer politically sensitive questions with less Chinese state framing. The weights, training data and evaluation materials are all published on Hugging Face, so anyone can inspect or rerun the work.

The recipe is deliberately plain supervised fine-tuning: about 52.8 million tokens over two epochs and 1,202 steps, with no reinforcement learning, no preference optimisation and no change to the architecture. A teacher model called Inkling generated 20,000 prompts spanning seven value-sensitive categories โ€” freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment.

On Nous's own 803-prompt test, judged blind by Grok 4.7, answers flagged as adopting the Chinese state's position fell from 22% on the base model to 0.9% on Arete-27B. The largest shift was on freedom of expression, where the flag rate fell from 60% to 1% while the factual score rose from 0.89 to 2.44. General capability stayed roughly flat: Arete scored 65.60% on Nous's Hermes Bench against 65.77% for the base model.

The headline numbers come from Nous itself, using its own scoring rules and an automated judge โ€” they have not been independently replicated. That is the point of publishing the eval: the 803 prompts, source materials, scoring instructions and training examples are all public, so other labs can test whether the shift holds under different rules.

Practically, the model is a 27-billion-parameter download of about 54 GB in BF16, and community quantisations let it run on consumer GPUs with 24 GB of VRAM. It is one of the first fully open values-realignment recipes built on a leading Chinese foundation model โ€” and a live test of how much of a model's politics is post-training, and how much is baked in from the start.

Sources