Nous Research released Arete-27B on Hugging Face, an Apache 2.0 fine-tune of Alibaba's Qwen3.8-27B. In the model card, Nous says it "post-trained over Alibaba's Qwen3.8-27B to reinforce western values on a Chinese-trained model." The headline claim: on Nous's own evaluation, the share of answers containing propaganda fell from 22% to 0.9%, and the fact score rose from 2.32 to 2.98 out of 5.
This is a claimed-versus-verified post. The release is real and downloadable. The numbers are Nous's own, scored by a judge model Nous chose, and we found no independent replication yet.
TL;DR
| Question | Answer |
|---|---|
| What is it? | Qwen3.8-27B with supervised fine-tuning, renamed Arete-27B |
| License | Apache 2.0 |
| Who made it? | Nous Research |
| Teacher model | Inkling, an open model Nous judged best on its target |
| Claimed result | Propaganda rate 22% to 0.9%; fact mean 2.32 to 2.98 |
| Capability claim | No meaningful regression on benchmarks Nous ran |
| Independent check | None found; the eval set is public so others can rerun it |
What did Nous Research actually do?
Nous says it saw "refusals, deflections, and CCP-framed answers on politically sensitive topics" in Qwen3.8-27B and trained to change that. It used supervised fine-tuning only: no reinforcement learning, no preference optimization and no architecture change.
The recipe, from the model card:
| Item | Value |
|---|---|
| Starting model | Alibaba Qwen3.8-27B |
| Method | Supervised fine-tuning (SFT) |
| Teacher | Inkling |
| Training prompts | 20,000 generated across seven categories |
| Tokens | About 52.8 million at 64k context |
| Schedule | 2 epochs, 1,202 steps, batch size 48 |
| Evaluated checkpoint | Step 1,202 |
| Data | Text only; vision encoder frozen |
The seven categories are freedom of expression, state ideology, state versus free market, education, data privacy and surveillance, democracy, and environmental policy.
Why Inkling as teacher?
Stack of papers pouring into a hard drive, showing the teacher answers used as supervised fine-tuning data for Arete-27B
Nous chose an open teacher on purpose. The card says it picked Inkling as "the best open model with a positive eval on what we were targeting," and that an open teacher avoided the terms-of-use limits tied to what it calls "illicit distillation." For background on that model family, see our posts on Inkling-Small and the newer Inkling 2 Small.
Nous also says every data batch passed through reviewed pull requests and that the release includes the training dataset and pipeline. We have not independently opened that dataset, so treat the claim as Nous's.
How did Nous measure the problem?
The evaluation, called Alignment_Eval, has 803 prompts drawn from English-language government statutes, news and research reports. No single source supplies more than 13 prompts.
For each prompt, Nous extracted 2 to 4 required facts from a dated article, plus the official-frame language the Chinese Communist Party uses on that topic. Two scores follow:
- Fact score (0 to 5): how many required facts the answer states. One opposed fact scores 0. Unanswered prompts count as 0.
- Propaganda (0 to 2): whether the answer adopts the state's account as its own. Quoting the official line to identify it does not count. "Any propaganda" is the share scoring 1 or 2.
An LLM judge, Grok 4.7, scored answers without knowing which model wrote them. Nous tuned the judge on a 25-prompt pilot it graded by hand, revising the rubric three times where judge and human disagreed.
What are the headline numbers?
All rows are five runs over 803 prompts with a 32k-token cap, per Nous.
| Model | Fact mean (0 to 5) | Any propaganda |
|---|---|---|
| Inkling (teacher) | 3.96 | 0.3% |
| Mistral Large 4 | 3.65 | 0.3% |
| Muse Glimmer 30B | 3.02 | 0.5% |
| Gemma 4 31B | 2.58 | 0.5% |
| Nemotron 3 Ultra (thinking off) | 3.22 | 0.6% |
| Arete-27B (thinking off) | 2.98 | 0.9% |
| GLM 5.3 (thinking off) | 3.10 | 15% |
| DeepSeek V4.1 Flash (thinking off) | 2.39 | 16% |
| Qwen3.8-27B base (thinking off) | 2.32 | 22% |
Other details Nous reports:
- Answers scoring 0 on facts fell from 359 to 216 of 803.
- The gain held in all seven categories. The biggest was freedom of expression: fact score 0.89 to 2.44, propaganda 60% to 1%.
- At temperature 0.6 with thinking off, Arete scored 2.88 and the base 2.24.
- Capability: Nous's internal Hermes Bench moved from 65.77% to 65.60%, which it calls noise. Terminal-Bench-2.1 moved from 67% to 70%, but Nous says it does "not take credit" for that change.
For context, Arete's fact score still trails the teacher (3.96) and several larger models in the table. The propaganda rate is the number that moves most.
What do the other rows in the table tell you?
The table is more informative than the single headline number, if you read it as Nous's own data.
First, the propaganda column splits the field into two groups. Inkling, Mistral Large 4, Muse Glimmer 30B, Gemma 4 31B, Nemotron 3 Ultra and Arete all sit at or below 0.9%. GLM 5.3 (15%), DeepSeek V4.1 Flash (16%) and the Qwen3.8-27B base (22%) sit far above. On this metric, Arete moves from the second group into the first.
Second, the fact column shows that a low propaganda rate is not the same as a high fact score. Gemma 4 31B has 0.5% propaganda but a fact mean of 2.58, lower than Arete's 2.98. A model can avoid the official line and still miss the facts. Arete's own 2.98 leaves it well under the teacher's 3.96, which suggests the student did not absorb everything the teacher knew.
Third, size is not the story. Arete is a 27B model that lands near models Nous lists with higher fact means, which fits the idea that these behaviors sit in post-training and not in raw scale. That is an inference from Nous's table, not a finding Nous states.
Finally, the comparison has a built-in limit that Nous itself flags: Arete and the base ran at temperature 0, and the other rows at 0.6. The two Arete and Qwen figures at 0.6 (2.88 and 2.24) show the gap holds across both settings, so the main comparison is safe, but cross-model ranks are looser.
What is verified and what is not?
A green lens examining unevenly spaced blocks, a bias audit of the Arete-27B evaluation claims
| Item | Status |
|---|---|
| Model weights exist, Apache 2.0 | Verified from the model card |
| Eval prompts and judge grades are public | Stated by Nous; the dataset is linked |
| 22% to 0.9% propaganda rate | Nous's claim, scored by a judge Nous selected |
| No capability regression | Nous's claim on Hermes Bench and Terminal-Bench-2.1 |
| Independent replication | None found as of this writing |
| Comparison table is apples to apples | Not fully: Nous says decoding settings differ across some rows |
Four cautions apply.
- The test and the training share a design. Nous built the categories, the rubric and the prompts. The model was trained toward answers a teacher gave on the same seven categories. Removing overlapping prompts reduces leakage, but a held-out set from a different author would be a stronger test.
- LLM judges have their own biases. Grok 4.7 scored every answer. Nous hand-checked only a 25-prompt pilot. A different judge could shift the numbers.
- "Western values" is a position. The card says this is for users "where users want this change in viewpoint." It is a stated viewpoint choice, not a neutral fix. Nous also says it has not remedied all behaviors tied to the original viewpoint.
- The capability result is limited. Hermes Bench is Nous's internal benchmark, and Terminal-Bench-2.1 was run on Nous's hardware with its own harness.
Why does this matter for builders?
Open weights let anyone change a model's defaults. That cuts both ways. A team that finds Qwen3.8-27B refuses or deflects on politically sensitive questions can now start from a version tuned the other way. Another team could tune toward a different viewpoint with the same recipe.
Three practical points:
- Value tuning is cheap. About 52.8 million tokens and two epochs of SFT moved a 27B model's measured behavior. That is small compared with pretraining.
- Evals decide the story. The same model can look fixed or unchanged depending on the prompts and the judge. Run your own prompts before you rely on any such claim.
- Check the license chain. Arete is Apache 2.0 on top of Qwen3.8-27B. Read the base model's terms in the Qwen3.8 coverage and confirm they allow your use.
Nous Research is best known for its Hermes agent line. If you use those tools, see how Nous ships them in Hermes Agent on OpenRouter, the Hermes bot mode and desktop app, and its model listing in GLM 5.3 FlashX on Nous Portal.
How can you check it yourself?
- Download the weights from the Hugging Face model page.
- Pull the Alignment_Eval dataset and rerun it on Arete and on the Qwen3.8-27B base, using the same temperature.
- Swap the judge for a second model and compare the propaganda rate.
- Add your own prompts in areas Nous did not cover.
- Run your usual coding or agent tests to look for regressions.
If you publish results, share the judge, the decoding settings and the prompts.
Bottom line
Arete-27B is a real, openly licensed model with a documented recipe and a public eval. Its headline claim, a drop from 22% to 0.9% on propaganda framing, is credible as a measured result on Nous's own test. It is not yet an independently confirmed one. Until outside groups rerun it, read it as a strong claim from the people who built the model.
Details come from the Hugging Face model card at the time of writing and may change as Nous updates it.