Is Arete-27B state of the art?Pending benchmarks.Announced Nov 6, 2026; no independent results yet.
Nous Research's open-weight 27B supervised fine-tune of Alibaba's Qwen3.8-27B. SFT only (~52.8M tokens, 64k context, two epochs / 1,202 steps); no RL, no preference optimization, and no architecture change. Nous frames the post-train as reinforcing Western values on a Chinese-trained base, with Inkling as the open teacher and the vision encoder frozen (text-only data). Weights on Hugging Face under Apache-2.0.
The details so far
- Lab
- Nous Research
- Announced
- Nov 6, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 27B
- Role
- General assistance and agentic work after value-alignment SFT; text generation
- Base
- Qwen/Qwen3.8-27B; separate from the Qwen3.8 27B catalog page
- Post-training
- Supervised fine-tuning only (~52.8M tokens, 64k context, 2 epochs / 1,202 steps). No RL, no DPO/preference optimization, no architectural modification (lab)
- Teacher
- Inkling (open teacher; lab)
- Frozen weights
- Vision encoder frozen; training data is text-only (lab)
- Framing (lab)
- Nous describes the goal as reinforcing Western values / altering CCP-framed refusals and answers on politically sensitive topics in the base model
- Checkpoints
- NousResearch/Arete-27B on Hugging Face; evaluated checkpoint step 1,202
- License
- Apache-2.0
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on Nous's 803-prompt propaganda evaluation, fact mean rose from 2.32 to 2.98 (out of 5) and answers containing propaganda fell from 22% to 0.9% (Arete-27B thinking off vs Qwen3.8-27B thinking off; model card / blog).
Lab claim: Hermes Bench (internal) 65.60% for Arete-27B vs 65.77% for Qwen3.8-27B — within about 0.2 points, which Nous treats as noise (blog / card).
Lab claim: Terminal-Bench-2.1 70% for Arete-27B vs 67% for the base under Nous's like-for-like run (blog); Nous does not take credit for the difference.
Sources
LABIntroducing Arete-27B (Nous Research)nousresearch.com/arete · Nov 6, 2026REPONousResearch/Arete-27B model card (Hugging Face)huggingface.co/NousResearch/Arete-27BLABNous Research on X: Arete-27Bx.com/NousResearch/status/2118822925445050435 · Nov 6, 2026Tracking continues until an independent board lists it.