Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values

Nous Research’s Arete-27B is a post-trained version of Alibaba’s Qwen3.8-27B, designed to reduce what the team describes as Chinese state-aligned political framing and make the model reflect Western values. Those values are assessed across seven topics, including free expression, democracy, privacy, markets, education, state ideology, and the environment. The team evaluated factual accuracy and whether answers adopted official Chinese government narratives, using 803 prompts tied to source material and a blinded language-model judge. On this test, the base model averaged 2.32 out of 5 for factual recall, and 22% of its answers showed propaganda by the evaluation’s definition. Nous selected Inkling as the teacher model and assembled training examples that did not overlap with the evaluation prompts.

The team fine-tuned all language-model weights on about 52.8 million tokens over two epochs, without reinforcement learning or preference optimization. Arete scored 2.98 for factual recall and reduced propaganda-containing answers to 1%; zero-scoring answers fell from 359 to 216. Improvements appeared in all seven categories, with the largest gains on freedom of expression. The model’s score on Hermes Bench was nearly unchanged from Qwen’s, while its Terminal-Bench score was higher, leading Nous to report no capability regression in the benchmarks it checked. The team argues these results show that a relatively small amount of targeted training can substantially change a model’s political framing without disrupting its underlying capabilities. It has released the weights, evaluation materials, and training data for further testing.