Illustrative image: a mountain arête — the cover of Nous Research's announcement, and the origin of the model's name. Image: Wikimedia Commons / Richard Webb, CC BY-SA 2.0 (Geograph).
📌 KEY HIGHLIGHTS
- Released 6 November 2026: Nous Research's Arete-27B is a supervised fine-tune of Alibaba's Qwen3.8-27B, not a model trained from scratch.
- The headline result: on Nous's Propaganda_Eval (803 prompts, seven value categories), answers containing state-aligned framing fell from 22% to 0.9%, and the fact score rose from 2.32 to 2.98 out of 5.
- No ability trade-off: Arete-27B scored 65.60% on Nous's Hermes Bench against 65.77% for the base model — a gap Nous calls indistinguishable from noise.
- Free to download: weights, training data and the test are public on Hugging Face under Apache 2.0. The full-precision weights are about 54 GB; smaller community versions fit a 24 GB consumer GPU.
- The catch: the test and its judge are Nous's own, and the model swaps one viewpoint for another — it is not viewpoint-free.
What happened?
Nous Research, the independent AI lab best known for its Hermes series of open models, announced Arete-27B on 6 November 2026. It took Alibaba's Qwen3.8-27B — one of the strongest open-weight models available, downloaded about 7.8 million times since Alibaba put it on Hugging Face in August — and post-trained it, in Nous's words, "to reinforce Western values on a Chinese-trained model".
The name does double duty: arete (ἀρετή) is the ancient Greek idea of excellence and virtue, and an arête is a sharp mountain ridge — the announcement's cover shows one. Nous chose Qwen3.8-27B because it was the strongest model it could run locally and because it showed the behaviour it wanted to change: in the lab's words, "refusals, deflections, and CCP-framed answers on politically sensitive topics". Rather than training from scratch, it rewrote the part that answers those questions.
Arete-27B Qwen propaganda explained: how Nous measured it
Before training, Nous built a test called Propaganda_Eval: 803 prompts across seven value categories — freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and the environment. Each prompt carries required facts and a source article, so answers can be scored for accuracy as well as framing.
An AI judge — Grok 4.7, from xAI — rated each answer on a 0–5 fact scale and flagged whether it adopted or endorsed the state's account. Merely quoting that account to identify it does not count; a flag needs an exact supporting phrase. The judge, prompts and scoring scripts are all published, so anyone can run the same test on any model.
The training itself was deliberately simple. Nous used supervised fine-tuning only — about 52.8 million tokens from roughly 20,000 generated questions, with answers supplied by a teacher model, Inkling from Thinking Machines Lab — over two epochs and 1,202 steps. There was no reinforcement learning, no preference optimisation and no change to the architecture; the vision encoder stayed frozen because the training data is text only.
| Model | Fact score (out of 5) | Answers flagged as propaganda |
|---|---|---|
| Qwen3.8-27B (base, thinking off) | 2.32 | 22% |
| Arete-27B (thinking off) | 2.98 | 0.9% |
| Arete-27B (thinking on, provisional) | 3.07 | 0.6% |
Source: Nous Research, "Introducing Arete-27B". Figures are Nous's own measurements on its own test.
The biggest shift came in freedom of expression, where the propaganda rate fell from 60% to 1% and the fact score rose from 0.89 to 2.44. Nous says the result lines up with the Western models it tested — and its punchline is that "the values in an open model are one fine-tune away".
Illustrative image: a neural network diagram. Arete-27B changes Qwen's behaviour with fine-tuning alone — no architecture changes. Image: Wikimedia Commons / Dake, Mysid, CC BY 1.0.
Why it matters for India readers
Open-weight models like Qwen are the workhorses of Indian developers, startups and researchers — free to download, free to run, and free to build on. That freedom has a catch this release makes visible: the values baked into a downloaded model travel with it. If you build a product on a model trained to deflect certain questions, your product deflects them too, until someone retrains it.
Arete-27B shows that retraining is cheap. Fifty-three million tokens is a rounding error next to the trillions behind a base model, yet it flipped the answers on hundreds of sensitive questions without denting general ability. That is good news for anyone who wants a Qwen-class model without the state framing — and a warning, because the same technique could push a model towards any viewpoint. A model de-biased this easily can be re-biased just as easily.
It connects to work already covered here: researchers edited Qwen's weights to fix its maths, and open-weight models are now small enough to run on your phone. The thread is the same — open weights let anyone inspect, test and change what a model does instead of taking the maker's word for it. Indian developers building on open models should still test, not assume, what theirs says on sensitive topics — and now there is a public test to run.
Should you care?
✅ Try it if you run open-weight models locally and want Qwen-level ability without the state-aligned framing: the weights are free on Hugging Face under Apache 2.0, and the test is public so you can check the claims yourself. ⏳ Wait if you need a hosted API or a model small enough for a laptop — there is no hosted inference yet, the full weights are about 54 GB, and the headline numbers come from Nous's own test.
FAQs
What is Arete-27B? Arete-27B is Nous Research's open-weight fine-tune of Alibaba's Qwen3.8-27B, released on 6 November 2026. It was retrained to reduce answers that adopt Chinese state positions on politically sensitive questions, and it is free to download under an Apache 2.0 licence.
How much does Arete-27B cost to run? The weights are free. There is no hosted API yet, so you run it yourself: the full-precision (BF16) download is about 54 GB, while community quantised versions start at 8.64 GB, with a 17.2 GB version sized for 24 GB consumer GPUs.
Does Arete-27B remove all bias? No. Nous says it reinforces Western values, not that the model is viewpoint-free, and it "does not claim to have removed every behaviour" tied to the base model's viewpoint. The 22%-to-0.9% figures also come from Nous's own test, so independent checks matter.
Can I run Arete-27B on my own computer? Yes, if you have a powerful GPU. Community GGUF versions run in llama.cpp, LM Studio and Ollama. Nous has said hosted inference is "locally for now", so there is no cloud version to try without downloading it.
Takeaway: Arete-27B proves a model's politics can be rewritten with one small fine-tune — useful if you want Qwen without the state framing, and a warning that any open model's values are only one fine-tune deep.
If you use open-weight models, run Nous's 803 prompts through the one you rely on — the test is public, and the answer might surprise you.
Sources
- Nous Research — Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values (primary source, published 6 November 2026)
- Hugging Face — NousResearch/Arete-27B model card (weights, training data and evaluation)
- MarkTechPost — Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9%
- Daily AI Brief — Nous Research releases Arete-27B, a Qwen model post-trained to remove CCP-aligned answers
- Sulat — Nous Research's Arete-27B retrains Qwen's politics with 53 million tokens
Comments
Post a Comment