Nous Research releases open-weight model Arete-27BMachine translation

Arete-27B is a supervised fine-tune of Qwen3.8-27B, with its weights, evaluation set and training data publicly available. In Nous Research’s own 803-prompt evaluation, the share of answers flagged for propaganda fell from 22% for the base model to 0.9%. The new model still scored zero on 216 prompts, and no hosted inference provider serves it yet.