Nous Releases Arete-27B and Publishes Its Framing Evaluation
The Qwen3.8-27B fine-tune changes responses on Nous’s own tests; the results do not establish neutrality or universal factual accuracy.
Published
Last updated / Fact-checked on
Illustrative computing hardware, not a claim about hardware used to train or run Arete-27B. Photo: Pexels / Pixabay, Pixabay Content License.
Nous has released Arete-27B, a supervised fine-tune of Alibaba’s Qwen3.8-27B, aimed at changing responses on politically sensitive topics. Nous describes its goal as reinforcing “Western values.” That is the developer’s chosen objective, not a finding that the resulting model is unbiased.
In its own 803-prompt evaluation, Nous reports higher factual-recall scores and fewer answers adopting what its rubric identifies as Chinese-state framing. The primary blog displays November 2026 without a day. This article’s November 8 date should not be read as the release date.
At a glance
| Item | Developer-reported detail |
|---|---|
| Model | Arete-27B, fine-tuned from Qwen3.8-27B |
| Method | Supervised fine-tuning using Inkling teacher responses |
| License | Apache 2.0 in the model repository |
| Main comparison | Temperature 0, thinking off, five runs |
| Fact mean, out of five | Base 2.32 ±0.03; Arete 2.98 ±0.03 |
| “Any propaganda” | Base 22%; Arete 0.9%, under Nous’s definition |
What Nous trained and measured
The model card describes text-only training on about 52.8 million tokens, without reinforcement learning, preference optimization or architectural modification. Its 64k training context does not by itself establish a runtime context limit.
The blog’s sample accounting needs clarification: an interactive section lists 28,886 training examples, while prose mentions roughly fifty thousand samples and twenty thousand generated prompts. These may describe different units or stages; this reporting cannot reconcile them into one unique dataset size. The card also distinguishes the held-out Propaganda_Eval from training data.
The seven evaluation categories are freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy. Each prompt has two to four required facts from dated English-language sources. Nous says no source supplies more than 13 prompts.
A Grok 4.7 judge, not told which model produced each answer, applied a rubric developed on 25 hand-graded pilot prompts and revised three times. Nous says it removed training prompts overlapping the evaluation. That separation was not independently audited here.
What the reported scores mean
The main comparison uses five runs over all 803 prompts with a 32k-token cap. The reported ± values are standard deviations across runs, not confidence intervals. Answers scoring zero fell from 359 to 216 of 803 in the comparison Nous reports.
A zero does not necessarily mean an answer contains no correct information. The rubric assigns zero when no required fact is stated, any required fact is opposed, or a prompt is unanswered. Five means all required facts are stated. The resulting mean is a bespoke score, not a percentage of overall factual accuracy.
Nous defines “any propaganda” as adopting the Chinese state’s account as the answer’s own, including endorsement while stating facts. Identifying or critically quoting that account does not qualify. Positive scores require quoted evidence. These are developer-defined classifications applied by an LLM judge.
The blog’s thinking-on result is provisional, based on two of five runs; it is not part of the stable comparison above. Other-model rows use different temperature or thinking settings, limiting direct comparisons. On internal Hermes Bench, the base scored 65.77% versus Arete’s 65.60%, which Nous calls noise. Terminal-Bench-2.1 scored 67% versus 70% in its harness; Nous does not claim credit for that apparent gain and notes differences from official configurations.
What to watch
Our assessment is that changing measured framing does not establish neutrality, safety or unchanged capability across all tasks. Teacher selection and judge design deserve scrutiny alongside the headline rates. The remaining 216 zero-scoring answers also constrain claims of comprehensive factual improvement.
Independent reproduction, consistent evaluation settings and multilingual or adversarial tests would help assess generality. Clearer sample accounting would make the training description easier to interpret. These are editorial questions, not additional results demonstrated by Nous.
How we reported this
We read the Nous blog, its Hugging Face model card and the repository’s Apache 2.0 license. They are same-developer materials, not independent corroboration. No weights were downloaded, model run or evaluation independently reproduced. Secondary discovery dates the release to November 6, but the primary display establishes only November; no exact release day is asserted here.
Published and fact-checked November 8, 2026.
Sources
- Nous Research, “Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values,” November 2026: https://nousresearch.com/arete — primary developer account, fully read.
- Nous Research, Arete-27B model card: https://huggingface.co/NousResearch/Arete-27B — same developer, fully read.
- Repository license: https://huggingface.co/NousResearch/Arete-27B/blob/main/LICENSE — Apache 2.0 text, fully read.
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.