Nous Research Releases Arete-27B, Cutting Qwen3.8-27B State-Aligned Propaganda From 22% to 0.9%
By Asif Razzaq Β· November 7, 2026
Nous Research has released Arete-27B, a supervised fine-tune (SFT) of Alibaba's Qwen3.8-27B. The model is a 27B open-weight release designed to strip Chinese state framing from Qwen's answers. On Nous's own 803-prompt evaluation, answers containing propaganda fell from 22% to 0.9%.
The pitch to developers is straightforward: retain Qwen-class agentic capability without state-aligned answers. The weights, eval set, and training data are all public.
TL;DR
- Size: 27B dense parameters (all active). Trained at 64k context; the base Qwen3.8-27B natively supports 262,144 tokens.
- Runs on: BF16 weights total 53.81 GB. A community Q4KM quantization weighs 17.20 GB, sized for 24 GB consumer GPUs.
- Performance: Capability was unchanged on the benchmarks Nous ran; political framing now aligns with Western models.
- Best result: Propaganda rate of 0.9%, down from 22% for base Qwen3.8-27B.
- Worst result: Fact score of 2.98 out of 5; 216 of 803 answers still score 0.
What Is Arete-27B?
Arete-27B is a supervised fine-tune of Qwen3.8-27B, a 27B dense open-weight model from Alibaba. Rather than a from-scratch training run, the release is a targeted behavioral intervention: Nous retrained the model's response tendencies to remove Chinese state framing while preserving the base model's reasoning and agentic skill.
The name references arete, the classical Greek concept of excellence or virtue β fitting for a model whose central claim is behavioral correction rather than raw capability gains.
The Evaluation: Propaganda Down, Capability Held
Nous evaluated the model against an in-house set of 803 prompts designed to surface state-aligned framing. The headline result is a drop in answers containing propaganda from 22% to 0.9% relative to the base model.
The team reports that capability was unchanged on the benchmarks it ran, positioning Arete-27B as a drop-in replacement for developers who want Qwen-class performance without the political framing.
The trade-off appears on factual grounding: Arete-27B scores 2.98 out of 5 on Nous's fact metric, and 216 of the 803 answers still score 0. That residual failure rate is the model's clearest limitation and a likely focus for follow-up releases.
Hardware and Deployment
Arete-27B was trained at a 64k context window, though the underlying Qwen3.8-27B supports up to 262,144 tokens natively. In BF16, the weights total 53.81 GB β practical on server-class GPUs but out of reach for most consumer hardware.
For local use, a community Q4KM quantization brings the footprint down to 17.20 GB, fitting within 24 GB consumer GPUs. As with any quantization, expect some quality degradation relative to the BF16 release.
Why It Matters in 2026
The release lands as the open-weight ecosystem increasingly treats alignment as an editable property rather than a fixed characteristic of a base model. Fine-tunes that reorient a model's political framing β while holding capability steady β give developers a middle path between adopting a base model wholesale and building alignment from scratch.
It also raises the governance question that has followed every de-framing release: if framing can be swapped at SFT cost, model provenance and behavioral guarantees become harder to audit. Nous's decision to publish the eval and training data alongside the weights is, in that light, as significant as the model itself.
Getting Started
- Weights: huggingface.co/NousResearch/Arete-27B
- Base model: huggingface.co/Qwen/Qwen3.8-27B
- Eval, training data, and details: nousresearch.com/arete
via MarkTechPost