Nous Research fine-tunes Qwen3.8-27B to move away from official PRC positions in responses

🤖 Nous Research fine-tunes Qwen3.8-27B to move away from official PRC positions in responses

Nous Research released Arete-27B — an open version of Alibaba's Qwen3.8-27B. The model was fine-tuned with SFT (~52.8 million tokens, no RL) on ~20,000 questions across seven topics, ranging from freedom of speech to ecology. The share of responses with CCP propaganda dropped from 22% to 0.9%, while Hermes Bench scored 65.60% compared to 65.77% for the base model.

🌍 The alignment of Chinese open models proved to be superficial — it can be removed through fine-tuning without losing agentic and coding capabilities (Terminal-Bench-2.1: 67% → 70%). Teams using such models locally can fine-tune them for their own jurisdiction without retraining from scratch.

👤 Running Qwen3.8-27B locally and encountering refusals on sensitive topics? Arete-27B is a ready-made replacement: same 27B parameters and 64k context, with weights available on Hugging Face.

Source 1: https://nousresearch.com/arete Source 2: https://huggingface.co/NousResearch/Arete-27B