Nous Research has released Arete-27B, an open-weight supervised fine-tune of Alibaba’s Qwen3.8-27B. It aims to remove Chinese state framing from the model’s answers on sensitive topics, MarkTechPost reports.
The details
- 27B dense, Apache 2.0, with weights, eval and training data all public.
- SFT only (no RL, no preference tuning), about 52.8M tokens at 64k context. The teacher was Inkling, an open model from Thinking Machines.
- On Nous’s 803-prompt eval, answers containing propaganda dropped from 22% to 0.9%, and the fact score rose from 2.32 to 2.98 out of 5.
- General capability held: Hermes Bench is within 0.2 points of base Qwen.
- BF16 is about 54GB. A community Q4_K_M GGUF is 17.2GB, which fits 24GB GPUs.
My take
Nous is upfront that it’s not perfect: 216 of the 803 answers still score zero on facts. Still, it’s a great example of what open weights let you do. I’ve got a plan to try it locally next to stock Qwen.
God Bless, and Use the Source!
