Nous research fine-tunes Qwen3.8-27B to significantly reduce state-aligned bias in arete-27b

Nous Research has introduced Arete-27B, an open-weight model derived from Alibaba’s Qwen3.8-27B through supervised fine-tuning. This modification aims to strip out Chinese state framing from the model's responses. Evaluation on Nous's 803-prompt test showed a dramatic reduction in propaganda-containing answers, dropping from 22% to just 0.9%. The fine-tuning process involved training on approximately 52.8 million tokens with a 64k context window over two epochs. The process utilized an open teacher model named Inkling from Thinking Machines, as Nous sought to bypass terms-of-use restrictions associated with other models like Nemotron 3 Ultra. The resulting Arete-27B maintains Qwen-class agentic capabilities while removing state alignment. Performance metrics showed improvements across the board; the fact mean increased from 2.32 to 2.98, and in the freedom of expression category, the fact score rose from 0.89 to 2.44. While the Hermes Bench score saw a minor dip from 65.77% to 65.60%, the Terminal-Bench-2.1 score improved from 67% to 70% within Nous's testing harness. The model weights are available on Hugging Face under the Apache 2.0 license.