❯ Nous releases Arete-27B, using about 52.8 million tokens to adapt Qwen3.8-27B’s response preferences

Adapting open weightsNous Research released Arete-27B, a post-trained version of Alibaba’s Qwen3.8-27B intended to adapt its value preferences and response behavior. The team also released model weights, training data and evaluation tools for reproduction and scrutiny. With roughly 27 billion parameters, the project demonstrates further customization of an existing foundation model.

Two supervised passesThe team trained for two epochs on about 52.8 million tokens using supervised fine-tuning with example answers, without reinforcement learning or architectural changes. Language-model weights were updated, the vision encoder was frozen and the multi-token prediction head was disabled. Teacher answers came from the open model Inkling, with training prompts checked for overlap against the evaluation set.

Maintaining task performanceNous reports 65.60% on its internal Hermes Bench, compared with 65.77% for base Qwen. Under the same local configuration, Terminal-Bench 2.1 scores were 70% and 67%, respectively, with the team interpreting the results as preservation of underlying capabilities and noting configuration differences as a possible explanation for local gains. The release also provides tools for examining changes in response preferences. Developers can obtain weights, data and the training description.

▮ SIGNALReleasing weights, training data and evaluations together lets users examine both the behavioral customization and the preservation of underlying task capabilities.