Nous Research Arete-27B: Open Qwen3.8 Fine-Tune Cuts Propaganda Answers

What Nous Research released

Nous Research has published Arete-27B on Hugging Face, a 27-billion-parameter text model built on Alibaba's Qwen3.8-27B and licensed under Apache 2.0. The weights are in BF16 safetensors, and community quantizations started appearing on the model page within hours. Because the base model is one of the most popular open models for local use, Arete is a drop-in option for anyone already running Qwen3.8-27B on their own GPU.

Nous describes the goal plainly: it post-trained the Chinese-trained model to "reinforce Western values," targeting refusals, deflections, and answers framed the way Chinese state media would frame them on politically sensitive questions. The model card lists seven value categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy.

How the model was trained

Arete is a supervised fine-tune only. Nous says it used no reinforcement learning, no preference optimization, and no architecture changes. The training run covered about 52.8 million tokens at a 64k context, over two epochs and 1,202 steps, with Inkling as the teacher model whose answers the fine-tune learned from. The vision encoder was frozen and the multi-token prediction head was disabled, so the change is confined to the language weights.

The blog post is less consistent on dataset size: it mentions twenty thousand prompts in one place, roughly fifty thousand samples in another, and 28,886 training examples in a chart caption. Readers who care about exact data volumes should treat those figures as unsettled until Nous clarifies.

The evaluation results

Nous built its own test, Propaganda_Eval, with 803 prompts across the seven categories, and published it on Hugging Face under an MIT tag. An LLM judge, Grok 4.7, scored each answer for factual quality from 0 to 5 and for propaganda from 0 to 2.

On that test, the share of base Qwen3.8-27B answers containing propaganda was 22%; Arete brought it to 0.9%, and the mean fact score rose from 2.32 to 2.98. The largest shift was on freedom of expression, where propaganda fell from 60% to 1%. For context, Nous's table puts GLM 5.3 at 15% and DeepSeek V4.1 Flash at 16% with thinking turned off, while Gemma 4 31B and Mistral Large 4 sit below 1%.

Capability looks intact. On Nous's internal Hermes Bench, Arete scored 65.60% against the base model's 65.77%, which Nous calls noise, and on Terminal-Bench-2.1 it scored 70% against 67%, an edge Nous says it does not take credit for.

Caveats before you switch

The headline numbers come from an evaluation that Nous designed, with a rubric Nous wrote and a single LLM judge, and no independent group has reproduced them yet. Decoding settings also differ across rows: Arete and the base model were scored at temperature 0, while other models ran at 0.6, where Arete's fact score drops to 2.88. Nous itself says it has not removed every behavior inherited from the base model.

For local users, the practical point is that Arete keeps Qwen3.8-27B's size, so the same hardware guidance applies: a quantized build fits on a single 24 GB GPU, while full BF16 weights need far more memory. If you run Qwen locally and want different behavior on political and civic questions, Arete is worth testing on your own prompts rather than taking the chart at face value.