MODEL
Arete-27B
modeltopic-notenous-researchopen-weightsqwen
Overview
Arete-27B is Nous Research‘s open-weight fine-tune of Alibaba‘s Qwen 3.8 27B base model, released with a 53M-token SFT run that Nous reports strips the model’s pro-Chinese-state behavioural bias from ~22% to <1% on sampled prompts while general capability metrics hold. Positioned as a usable debiased checkpoint for standard open-weights inference workloads — behavioural de-biasing, not weight-level removal of ideological structure.
Timeline
- 2026-11-08-AI-Digest — Nous Research fine-tuned Alibaba‘s open-weight Qwen 3.8 27B into Arete-27B using a
53M-token SFT run and reports pro-Chinese-state behavioural bias dropping from~22%to<1%with general capability metrics held. Nous’s own framing — that the bias “sat in a thin layer late in training” and fine-tuning can strip it — informs the headline. Nuance worth carrying: behavioural de-biasing on sampled prompts ≠ weight-level removal; the<1%figure measures sampled prompt behaviour, not the removal of ideological structure from the weights. The Decoder itself notes fine-tuning “does not necessarily remove biases deeply embedded during original training,” and the Nous result has not been adversarially red-teamed at scale against prompt-smuggling, activation steering, or system-prompt bypass. Prior claims in this shape (Hirundo89.8%→2.8%, Lazarus AI “ReAligned” LoRA) have not all survived adversarial follow-up. For practitioners, the practical consequence is: Arete-27Bis a usable debiased checkpoint today for standard inference workloads, but shipping it into safety-critical or adversarial contexts without independent red-team eval would be premature. Log against MOC - Open Source Models.
Key Developments
53M-Token SFT De-Biasing of Qwen 3.8 27B (November 2026): Reported drop from~22%to<1%pro-state behavioural bias on sampled prompts with general capability metrics held. Behavioural de-bias, not weight-level removal — the practitioner read is “usable debiased checkpoint for standard inference, not yet adversarially red-teamed”.
Related
See also: Nous Research, Qwen 3.8 27B, Alibaba, MOC - Open Source Models.