Reported News
Nous Research releases Arete-27B, a Qwen3.8-27B fine-tune that cuts propaganda answers to 0.9%
Nous Research released Arete-27B on Hugging Face on Nov 6, 2026, an Apache 2.0 supervised fine-tune of Qwen3.8-27B. The model card says that on its 803-prompt evaluation it "raised its fact score from 2.32 to 2.98 out of 5 and cut answers containing propaganda from 22% to 0.9%."
- Changes reported
- 5
- listed below, each from a source we read
- We tested
- Not tested yet
- reported from sources; we did not run it ourselves
- Gaps we flag
- 3
- things we could not verify
- Sources
- 2
- 1 primary, linked below
What changed
| # | What the source says | Status |
|---|---|---|
| 01 | Arete-27B is a 27B-parameter post-train of Alibaba's Qwen3.8-27B using supervised fine-tuning only, with no RL, preference optimization or architecture change. The card lists Inkling as the teacher model, text-only data and a 64k training context. | Reported |
| 02 | Training used about 20,000 prompts across seven value categories, such as freedom of expression, democracy and data privacy, and about 52.8 million tokens over two epochs and 1,202 steps, according to the model card. | Reported |
| 03 | On the 803-prompt propaganda evaluation, graded by an LLM judge, the card reports the fact mean rising from 2.32 to 2.98 out of 5 and answers with propaganda falling from 22% to 0.9%; freedom of expression moved most, from a 60% to a 1% propaganda rate. | Reported |
| 04 | General capability held in Nous's runs: Hermes Bench 65.60% versus 65.77% for Qwen3.8-27B, which the card calls noise, and Terminal-Bench-2.1 70% versus 67%, an improvement Nous says it does not take credit for. | Reported |
| 05 | Weights are on Hugging Face under Apache 2.0, and the evaluation set with all 803 prompts and source articles is published as a separate dataset; the card points to the blog post for the training dataset release. | Reported |
reported by the sourcemeasured or run by usnot verified
Primary source: Arete-27B model card (Hugging Face). We are the fast, independent index; the vendor’s own page is the source of record for this release.
In the source’s words
“We used supervised fine-tuning only. No RL, no preference optimization, and no architectural modification.”
“On Hermes Bench, the scores are within 0.2 percentage points, which we interpret as noise.”
Early verdict
Of interest to teams that want an open 27B model with Qwen3.8-27B's capability but less state-aligned framing on political topics; treat the gains as the vendor's measurement until others rerun the published eval.
Basis: Based only on the model card and evaluation dataset page as read on Nov 8, 2026; we have not run the model.
What we have not verified
- We did not run Arete-27B or the evaluation; all scores are Nous Research's own, graded by an LLM judge on Nous's own benchmark.
- The card notes Arete-27B and Qwen3.8-27B ran at temperature 0 while other models in its table used 0.6, so cross-model rows are not like-for-like.
- We did not check the training dataset release, which the card says is announced in the blog post.
This brief is Reported: written from the sources below, not from running the release ourselves.
Sources
- Arete-27B model card (Hugging Face)primary source · fetched 2026-11-08
- Propaganda_Eval dataset (Hugging Face)source · fetched 2026-11-08
Related
Spotted an error? See our corrections policy or email the address on the contact page. Markdown version for agents: /news/nous-arete-27b.md
Want these briefs in your Google Top Stories? Add Agentic Weekly as a preferred source on Google.
Newsletter
The Weekly Verdict. Free, every Monday.
A free email every Monday: the week's few big AI releases, then 5 short use cases for the new models, each with a prompt you can paste.
Free. Double opt-in. One click to unsubscribe.