Arete 27B
Nous post-trained Alibaba's Qwen3.8-27B using text from an Inkling teacher. It frames the work as a study of how post-training changes model values.
Training and model behavior
The supervised training used about 52.8 million tokens and seven topic categories. It updated language-model weights while freezing the vision encoder. Training used a 64K sequence length. The published evaluation measures the developer's chosen definition of propaganda, alongside general capability checks.Arete 27B model card
Working with this model
Practical guidance. Read the evaluation prompts and scoring rules before interpreting alignment scores. Compare answers against independent sources and include viewpoints beyond those in the training categories.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | NousResearch/Arete-27B |
| Context window | 262,144 tokens Configured position limit; practical serving limits depend on the runtime. Arete 27B configuration |
| Base model | Qwen/Qwen3.8-27BArete 27B model card |
| Repository license label | apache-2.0Arete 27B model card |
| Parameter scale | 27BArete 27B model card |
| Training method | Supervised fine-tuning; no reinforcement learning or architecture change.Arete 27B model card |