Arete-27B cuts propaganda responses in Qwen3.8-27B fine-tune

Nous Research has released Arete-27B, a 27B dense open-weight model created by supervised fine-tuning Alibaba’s Qwen3.8-27B. The model keeps the base architecture and Apache 2.0 license. The training used no reinforcement learning or preference optimization, and the vision encoder remained frozen because the dataset was text-only.

On Nous Research’s 803-prompt evaluation, answers judged to adopt a state account fell from 22% for Qwen3.8-27B to 0.9% for Arete-27B. The average fact score increased from 2.32 to 2.98 out of five, although 216 of 803 answers still received a zero. In the freedom-of-expression category, the fact score rose from 0.89 to 2.44 and the propaganda rate declined from 60% to 1%.

Capability results were broadly unchanged on the reported tests. Hermes Bench scored 65.60%, compared with 65.77% for the base model, while Terminal-Bench-2.1 reached 70% in Nous’s evaluation harness. The training run used about 52.8 million tokens and produced a public dataset of 34,878 prompts. BF16 weights require 53.81 GB; a community Q4_K_M version is 17.20 GB. The evaluation set is MIT licensed, and no hosted inference provider currently serves the model.