Nous Research has released Arete-27B, an open fine-tune of Alibaba's Qwen3.8-27B that it says cut propaganda-tagged answers from 22% to 0.9% on its own 803-prompt test. The model card states the goal plainly: "We post-trained over Alibaba's Qwen3.8-27B to reinforce western values on a Chinese-trained model." The release is unusual for saying out loud what retuning a model's politics means, and it ships under the permissive Apache 2.0 licence.
The method was deliberately simple. Nous used supervised fine-tuning only, with no reinforcement learning, no preference optimization and no change to the model's architecture. Training ran about 52.8 million tokens over two epochs and 1,202 steps at a 64,000-token context. A separate open model called Inkling acted as "teacher," generating 20,000 prompts across seven categories including freedom of expression, state ideology, data privacy and surveillance, democracy and environmental policy.
Grading leaned on other models. A single LLM judge, Grok 4.7, scored each answer against a fixed rubric and a set of required facts, without being told which model wrote it. On a zero-to-five factual-recall scale drawn from English-language statutes, news and research, base Qwen scored 2.32, Arete 2.98 and the Inkling teacher 3.96. General capability barely moved: Hermes Bench went from 65.77% to 65.60%, which Nous calls noise, and Terminal-Bench-2.1 rose from 67% to 70%.
Those numbers deserve caution. The headline drop from 22% to 0.9% is measured on Nous's own eval, graded by one model against a rubric Nous wrote, so it shows the fine-tune moved answers toward the values Nous selected, not that it made them more accurate in any neutral sense. The categories, from "state versus free market" to "environmental policy," are contested political questions, and "reinforce western values" is a point of view, not a correction. The factual-recall gain from 2.32 to 2.98 is the more checkable result, and even there Arete trails its teacher.
The release lands on a base model whose own successor is imminent. Alibaba says Qwen 4 is "currently in training," and its coding tools were updated in late October to recognize Qwen 4 identifiers, though no model card, price or licence has appeared; a remark attributed to Alibaba's CTO pointed to a release within days. Arete is built on the current Qwen3.8-27B, which ships under Apache 2.0, the most permissive of the three licences in that generation and the reason Nous could redistribute a modified version freely.
Arete is also the latest in a steady stream of Western open-weights work built on Chinese base models, which Nous itself notes. That reflects a practical reality: several of the strongest open models now come from Alibaba, DeepSeek and others, and developers who want to change what those models say must start from weights someone else trained. "The users of a model deserve optionality in which values they work with," Nous writes, a framing that is itself a claim worth examining as value-tuned forks multiply.