Nous Research releases Arete-27B fine-tune

Nous Research announced Arete-27B in materials dated Nov. 6, 2026, an open-weight, 27-billion-parameter supervised fine-tune of Alibaba’s Qwen3.8-27B rather than a model trained from scratch. The company says it post-trained the model to reinforce what it describes as Western values and released the weights, evaluation and training data publicly.

What did Nous Research release?

Nous Research released Arete-27B as a public, open-weight model and described it as a supervised fine-tune of Qwen3.8-27B. The model card identifies Nous Research as the developer, Qwen3.8-27B as the starting model, and supervised fine-tuning as the post-training method.

The company said the release package included the model weights, the evaluation set and the training data, and its announcement says the repository on Hugging Face is public rather than gated. The listed license is Apache 2.0.

How was Arete-27B trained?

Nous Research says it chose Qwen3.8-27B because it was the most performant local model and already showed the behavior it wanted to change. In its announcement, the company said it observed refusals, deflections and CCP-framed answers on politically sensitive topics.

The model card says Arete-27B was trained using supervised fine-tuning only, without reinforcement learning, preference optimization or architectural changes. Nous Research says it used Inkling as the teacher model after selecting it as the best open model with a positive evaluation on the target behavior, and that using an open teacher helped avoid terms-of-use restrictions associated with illicit distillation.

Training used about 52.8 million tokens at 64k context over two epochs and 1,202 steps, with no sequence packing. The model card lists AdamW, a batch size of 48, bf16 compute with fp32 master weights, tensor parallelism of 4, and frozen vision-encoder weights. Multi-token prediction was disabled.

Nous Research also says it generated 20,000 prompts across seven categories and collected teacher responses alongside the base model’s own. The company says each batch of training data went through reviewed pull requests and that it has published the dataset.

What did Nous Research test?

The company’s main evaluation used 803 prompts across seven value categories: freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment. Nous Research says the prompts came from English-language government statutes, news and research reports, and that no single source supplied more than 13 prompts.

For each prompt, Nous Research says it extracted two to four required facts from a dated source article and added official-frame language used by the Chinese Communist Party on the topic. The company says it built the judge on a 25-prompt pilot graded by hand and revised the rubric three times, and that it removed any training prompts that overlapped with the evaluation set before training.

These evaluation, methodology and benchmark results come from Nous Research’s announcement and model card, and Venture Herald has not independently reproduced them. On Nous Research’s evaluation, the model card says factual recall improved from 2.32 to 2.98 out of 5 and answers containing propaganda fell from 22% to 0.9%. The model card also says the number of zero-scoring answers fell from 359 to 216 out of 803. Nous Research’s X post rounded the propaganda figure to 1%.

According to Nous Research’s model card, the reported improvement held across all seven categories. It says the biggest shift came in freedom of expression, where the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%.

Did the benchmark results change model capability?

Nous Research says its benchmark runs did not show a meaningful capability regression, though it also notes that the benchmark gap on Hermes Bench was small enough to treat as noise. The model card reports 65.60% for Arete-27B versus 65.77% for Qwen3.8-27B on Hermes Bench, and 70% versus 67% on Terminal-Bench-2.1.

The model card says the reported rows were not run under identical decoding settings across all models. It also warns that the linked Propaganda_Eval dataset is the held-out evaluation set, not the training set.

Nous Research says the model is intended for general assistance and agentic work for users who want this change in viewpoint, and for research into how post-training affects values and behavior. The company frames the project as evidence that values in an open model can be changed with a relatively small fine-tune.