Nous Research released Arete-27B, an Apache 2.0 open-weight model derived from Alibaba’s Qwen3.8-27B. The team fine-tuned the base model on approximately 53 million tokens of supervised data. The weights and training data for Arete-27B are available under an Apache 2.0 license on Hugging Face.
The model reduced its propaganda rate from 22 percent to 0.9 percent on Nous’s internal evaluation set. xAI’s Grok 4.7 served as the judge model to identify propagandistic answers. The evaluation dataset consisted of 803 prompts across seven categories including free speech and state ideology.
General capabilities remained largely intact, with the Hermes Bench score staying flat at 65.60 percent. Nous used the Inkling model from Thinking Machines Lab as the teacher for the fine-tuning process. Developers can now inspect and modify the code to align outputs with specific safety requirements.
