Nuss Research Releases Arete-27B, Cutting Propaganda-Style Answers to 0.9%
The fine-tune of Alibaba's Qwen3.8-27B was published on Hugging Face under Apache 2.0, with its weights, training data and evaluations.
Nuss Research released Arete-27B, a supervised fine-tune of Alibaba's open-weight Qwen3.8-27B, according to aitimes.com. The group published the model's weights, training data and evaluation materials on Hugging Face under the Apache 2.0 license on November 6, 2026.
Nuss Research reports that supervised fine-tuning alone lowered the share of pro-Chinese-government propaganda-style answers from 22% to 0.9%. In the model's own evaluation, its fact score rose from 2.32 to 2.98 out of 5.
The shift was larger in the free-expression category: the fact score there rose from 0.89 to 2.44, and propaganda-style answers fell from 60% to 1%. Answers scoring zero on facts dropped from 359 to 216.
Gaps remain. About 27% of the 803 questions still drew answers that omitted needed facts or contradicted them.
Teacher model and benchmark results
Training used Thinking Machines' open model Inkling as the teacher, with about 52.8 million tokens of data. On the internal Hermes Bench, Arete-27B scored 65.60% against 65.77% for the original model; on Terminal Bench 2.1 it reached 70% versus 67%.
Developers have commonly produced uncensored versions of Chinese open-weight models such as Qwen and DeepSeek when they are released.
Quick answers
What is Arete-27B built on?
It is a supervised fine-tune of Alibaba's open-weight Qwen3.8-27B.
Where can I get Arete-27B?
Its weights, training data and evaluation materials were published on Hugging Face under the Apache 2.0 license.
How did Arete-27B score on benchmarks?
It scored 65.60% on the internal Hermes Bench versus 65.77% for the original, and 70% on Terminal Bench 2.1 versus 67%.