Nous Research released Arete-27B on November 6, 2026, and the headline number is hard to ignore: answers that adopted the Chinese state’s account of events fell from 22% in the base model to 0.9% in Arete, according to Nous’s own evaluation. Arete is Alibaba’s open-weights Qwen3.8-27B, post-trained by Nous “to reinforce Western values on a Chinese-trained model.” The model did not get bigger or smarter in any general sense. It got retrained on what to say when politics comes up — and on what facts to include when it answers.
If you build client content, tools, or chatbots on open models, that distinction matters more than the benchmark table. You can swap a system prompt in five minutes. You cannot prompt away a bias that lives in the weights. Arete is a public experiment in whether a small, cheap fine-tune can move that bias — and a warning that the same trick can install any values at all. Every number in this article is self-reported by Nous, which built the test, picked the judge, and ran the comparisons. No independent lab has rerun the evaluation yet.
What Arete-27B Actually Is
Arete-27B is not a new model family. It is Qwen3.8-27B, released by Alibaba in August 2026, put through one more round of training by Nous Research. Qwen3.8-27B is a 27-billion-parameter dense model with a 262,000-token context window. In BF16 precision it takes up roughly 54 GB, so it runs on a serious single GPU or a small multi-GPU setup, not on a laptop. The weights are published on Hugging Face under the Apache 2.0 license, so anyone can download them, run them, and fine-tune them further without paying Nous or Alibaba a cent.
That openness is the point. Closed models already come with their own baked-in political guardrails, and you mostly have to take the provider’s word for what those guardrails are. With an open-weights model, the training is visible and the weights are downloadable — and in this case the evaluation set is public too. Nous published the training data (arete-data-pairs) and the evaluation (Propaganda_Eval) alongside the weights.
“Arete” is a Greek word for excellence or virtue, and Nous is framing this as a values project, not a capability project. The model was not trained to write better code or solve harder math. It was trained to answer political questions the way a Western-trained model would, without repeating the Chinese state’s account of contested events. Whose “Western values,” exactly? The answers in the training data came from a teacher model, Thinking Machines’ Inkling, and the evaluation judge was xAI’s Grok 4.7. So “Western values” here means, in practice, the values and factual framing that Inkling produces and Grok 4.7 rewards. Every values fine-tune has a teacher, and the teacher’s blind spots become the student’s blind spots.
How Nous Retrained Qwen’s Politics
The method is surprisingly simple. The entire intervention was supervised fine-tuning — SFT — and nothing else. In Nous’s words: “No RL, no preference optimisation, and no architectural modification.” No reinforcement learning. No preference ranking. No changes to the model architecture. Just examples of good answers, shown to the model until it learned the pattern.
The scale is small for a 27B model: about 52.8 million tokens of training data, 2 epochs, 1,202 steps, a 64,000-token training context, roughly 20,000 prompts and 28,886 training examples. Full pretraining runs measure data in trillions of tokens. Arete’s retraining used tens of millions. If the headline results hold up, a small, cheap intervention moved a big model’s political behavior a lot.
The teacher was Thinking Machines’ Inkling, which generated the target answers the student was trained to imitate. The vision encoder was frozen and the training data was text-only, so this was a language-behavior change, not a multimodal one. The model learned which facts to include, how to frame contested topics, and when to push back against a loaded premise instead of going along with it.
That “SFT only” detail sets both a ceiling and a floor. It means the technique is cheap and reproducible — any competent team with GPUs and a teacher model can do it. But SFT is imitation. The model learns to produce answers that look like the teacher’s. That does not guarantee the change holds up under adversarial prompting, jailbreaks, or topics outside the training distribution. Nous is not claiming otherwise. Its model card says: “That is not to claim that we remedied all behaviors associated with the extant viewpoint.” That is the most honest sentence in the release.
The Evaluation: 803 Prompts, 7 Categories, One AI Judge
To measure whether the retraining worked, Nous built its own evaluation, Propaganda_Eval, and published it. The eval has 803 prompts across seven categories: freedom of expression, state ideology, state vs free market, education, data privacy and surveillance, democracy, and environment. Each prompt was built from a dated source — a statute, a news report, or a research paper — and came with two to four required facts the answer was supposed to include.
The judge was xAI’s Grok 4.7. It graded blind, never told which model wrote the answer it was scoring. It scored factuality from 0 to 5 based on how many required facts appeared, and it flagged answers that adopted the Chinese state’s account. That flag is the “propaganda rate” in the results below: the share of answers that repeated the state line.
The setup has real strengths. The prompts are grounded in dated sources, so the required facts are checkable. Blind grading removes one obvious source of bias. Publishing the eval, the judge prompt, and the scoring scripts means anyone can rerun it. But there is a limit you cannot wave away: the judge is a single large language model with its own leanings. Swap Grok 4.7 for a different judge and the numbers would move. How much is unknown, because nobody has done that swap publicly yet.
The Results Nous Reported
Every number in this section is self-reported by Nous. The headline: answers adopting the Chinese state’s account fell from 22% in the base Qwen3.8-27B to 0.9% in Arete with thinking mode off, and 0.6% with thinking mode on. The average fact score rose from 2.32 to 2.98 out of 5. Arete did not just stop repeating the state line — it also included more of the required facts. That combination matters. A model that dodges a question entirely would score zero on propaganda and zero on facts. Arete, by Nous’s numbers, answered more fully and more factually.
The biggest shift came in freedom of expression, where the base model was weakest. There, the fact score jumped from 0.89 to 2.44, and the propaganda rate fell from 60% to 1%. If you have ever watched a Chinese-trained open model hedge, deflect, or recite an official line on a free-speech question, you know what a 60% rate looks like in practice. A 1% rate, if it holds, means the behavior is mostly gone on the tested prompts.
Nous also ran other models through the same eval. The Inkling teacher scored 0.3% propaganda. Mistral Large 4 scored 0.3%, and Gemma 4 31B scored 0.5%. The three Chinese-developed models in Nous’s table landed between 15% and 22%. So Arete at 0.9% sits close to the Western-developed models and far from where its own base model started.
Model-by-model results on Propaganda_Eval (as reported by Nous)
| Model | Propaganda rate | Fact score (out of 5) | Notes |
|---|---|---|---|
| Qwen3.8-27B (base) | 22% | 2.32 | Alibaba’s model before retraining |
| Arete-27B (thinking off) | 0.9% | 2.98 | Nous’s fine-tuned version |
| Arete-27B (thinking on) | 0.6% | Not separately reported | Same model, thinking mode enabled |
| Inkling (teacher) | 0.3% | Not separately highlighted | Thinking Machines’ model; generated the training answers |
| Mistral Large 4 | 0.3% | Higher than Arete | Western-developed comparison model |
| Gemma 4 31B | 0.5% | Higher than Arete | Google’s open model in the comparison |
| Chinese-developed models (3) | 15–22% | Lower than Arete | Range across the three models in Nous’s table |
Read the table carefully: Arete’s 2.98 fact score still trails Mistral Large 4, Nemotron 3 Ultra, and GLM 5.3. The propaganda rate fell dramatically, but factual completeness did not leap to the top of the pack. Arete answers more factually than its base model did, and it still leaves facts on the table compared with the strongest models in the comparison.
What Did Not Change: General Capabilities Stayed Flat
If a fine-tune rewires a model’s politics, a fair worry is that it breaks something else. Nous checked two capability benchmarks, and the results are reassuringly boring. On Terminal-Bench 2.1, a test of agentic coding and terminal tasks, Arete rose from 67% to 70%. On Hermes Bench, Nous’s own general benchmark, it moved from 65.77% to 65.60% — essentially unchanged.
The Honest Limits: What Arete Does Not Prove
Read this section before you repeat any number from this article to a client. Everything above is self-reported. Nous built the evaluation, picked the judge, chose the comparison models, and ran the tests. That does not mean the numbers are wrong. It means they have not been independently verified, and independent verification is the step that turns a vendor claim into a fact you can rely on.
Start with the judge. A single LLM — Grok 4.7 — graded every answer. Blind grading helps, but a judge is not a neutral measuring instrument. A different judge could score the same answers differently, and nobody has published that comparison yet. Until someone reruns Propaganda_Eval with a second judge, or with human graders, the 0.9% figure is best read as “0.9% according to Grok 4.7 on Nous’s prompts.”
Then there is the decoding setup. Arete and the base Qwen model ran at temperature 0, the most deterministic setting. The comparison models ran at temperature 0.6, which introduces randomness. That is not an apples-to-apples comparison. It may not explain a gap from 22% to 0.9%, but it muddies the comparison-model numbers, and an honest write-up should say so plainly.
Look closer at the fact scores and the picture gets more mixed. Even after retraining, 216 of the 803 answers still scored zero on factuality — more than a quarter of the test set. The average fact score of 2.98 out of 5 is better than the base model’s 2.32, but it is not a high score in absolute terms. The typical answer still missed required facts. Arete is less likely to repeat the state line, and it is still often incomplete. Both things are true at once.
Finally, do not present Arete as neutral. “Neutral” implies the model has no viewpoint. Every model trained on human text has a viewpoint, and every fine-tune chooses one. Arete was trained to reinforce Western values. That is a viewpoint, stated openly by its creators. It may be the viewpoint you prefer — for many US-facing tools, it probably is — but calling it neutral hides the choice that was made. Nous itself does not claim neutrality. It claims a deliberate shift from one set of values to another, and it warns that it did not remedy every behavior tied to the old viewpoint.
Why This Matters If You Build Client Tools on Open Models
If you build chatbots, content tools, research assistants, or customer-facing AI features for clients, the base model you pick is a product decision, not just a technical one. Your client will not see the weights. They will see the answers. And if those answers repeat a foreign government’s line on a sensitive topic, that is your reputation on the line, not Alibaba’s.
The bias lives in the weights, not just the prompt. You can write a careful system prompt that says “answer neutrally, include multiple perspectives, do not repeat state propaganda.” That helps. It does not rewrite what the model learned during pretraining. Under pressure — a leading question, a jailbreak attempt, a topic phrased in a way the system prompt did not anticipate — the base model’s instincts can show through. A fine-tune changes those instincts. A system prompt polices them. Those are different levels of intervention, and for a client-facing tool they are not interchangeable.
Arete does not eliminate that risk. A 0.9% self-reported rate on one eval, graded by one judge, is not a guarantee. But it changes the risk profile in a way a system prompt alone does not. If you are shipping a tool where political questions are in scope — and on the open internet, they always are, because users ask whatever they want — then the choice of base model deserves the same care as the choice of hosting or UI. The same logic applies beyond chat models: if your client tool transcribes audio, NVIDIA’s open Nemotron transcribe model deserves the same kind of evaluation before you build on it.
Arete vs Plain Qwen3.8 vs a US Open Model: Which Call Is Right?
There is no single right answer. The right call depends on what you are building, who will use it, and how much political content the tool will touch.
Choose Arete-27B when you want Qwen3.8’s capabilities — its 262K context window, its general performance, its open license — but you need its political answers steered toward Western framing for a US audience. This is the natural pick for a US-facing content or research tool built on Qwen. The cost is that you are trusting Nous’s self-reported eval until an independent rerun lands, so test the model yourself on your own prompts before you ship it.
Choose plain Qwen3.8-27B when political framing is not in scope for your tool, or when you plan to do your own fine-tune for a specific domain. If your tool only writes product descriptions, summarizes internal documents, or generates code, the political behavior may never surface. But be honest with yourself about “not in scope.” If users can type free text, political questions are in scope, whether you planned for them or not.
Choose a US open model — Mistral Large 4 or Gemma 4 31B are the two in Nous’s own comparison — when you want the lowest propaganda rate in the published table and you do not need Qwen specifically. Mistral Large 4 scored 0.3% with a higher fact score than Arete. Gemma 4 31B scored 0.5%, also with a higher fact score. The trade-off is capability, licensing, context length, and ecosystem: you are switching model families, which means retesting prompts and revalidating whatever you built on Qwen. For some projects that is worth it. For others, Arete’s in-family fix is the cheaper path. If you want another open-weights option to compare against, Kimi K3.1 Pro’s weights are live with a permissive license — worth a look before you commit to one model family.
Try It Yourself: Download Arete and Rerun the Eval
You do not need to take Nous’s word for any of this. The weights, the data, and the eval are public. First, find the model on Hugging Face under Nous Research’s organization page. The weights are under Apache 2.0, so you can download and use them freely, including commercially, subject to the license terms. Budget about 54 GB of disk for the BF16 weights, and make sure you have a GPU setup that can actually load a 27B model — this is not a model you run on a laptop without quantization.
Second, pull the evaluation set. Propaganda_Eval is published alongside the model, and it includes the 803 prompts, the required facts for each, the judge prompt, and the scoring scripts. Run the base Qwen3.8-27B and Arete-27B through the same prompts, with the same decoding settings, and compare the outputs yourself. Use temperature 0 for both, so your comparison is cleaner than the published one.
Third, add your own prompts. The published eval covers seven categories, but your users will ask about topics outside those categories. Write twenty or thirty prompts that match your actual use case — the questions your client’s customers ask, the topics your content tool covers, the edge cases you worry about. Run both models on those prompts and read the answers. This is the test that tells you whether Arete helps your tool, not just whether it passes Nous’s test.
Fourth, test for breakage, not just for politics. Run your normal prompts — coding, writing, summarization, the tasks your tool actually performs — and confirm Arete handles them as well as the base model. Nous’s capability benchmarks say it should, but your prompts are the ones that count.
You can read Nous’s full write-up at the official Arete page, and a clear independent summary of the release at FourWeekMBA’s coverage. Start with the official page for the method and the numbers, then use the summary to sanity-check your understanding. Both are worth ten minutes before you download 54 GB of weights.
The Flip Side: “One Fine-Tune Away” Cuts Both Ways
Nous’s most quoted line from this release is: “The values in an open model are one fine-tune away.” It is a good line because it is true, and it is a worrying line for the same reason. If a small SFT run can move a model from 22% state-line answers to 0.9%, then a small SFT run can move a model in the other direction just as easily. The same technique that installed Western values in Arete can install any values at all — a different political line, a commercial agenda, a set of “facts” that are not facts.
That is not a reason to dismiss Arete. It is a reason to understand what open models actually are. An open-weights model is not a neutral artifact. It is a starting point. Whoever fine-tunes it last decides what it says when the hard questions come up. That is true whether the fine-tuner is Nous Research, a startup, a government, or you. The weights do not remember who trained them first. They reflect who trained them last.
Three things will tell you whether Arete’s numbers hold up. First, an independent rerun of Propaganda_Eval — ideally by a lab with no connection to Nous, using a different judge, with human grading on a sample. Second, adversarial testing: prompts designed to break the fine-tune, jailbreaks, and topics outside the seven tested categories. SFT is imitation, and imitation can be brittle. Third, adoption: whether developers actually switch to Arete for US-facing tools and publish their own comparisons. Until then, the fair summary is this: Nous did a cheap, transparent, well-documented fine-tune that, by its own measurement, cut state-line answers from 22% to 0.9% while leaving general capability flat. The method is real, the data is public, and the limits are serious. Download it, test it on your prompts, and decide for yourself.
Frequently Asked Questions
What is Arete-27B?
Arete-27B is Nous Research’s post-trained version of Alibaba’s open-weights Qwen3.8-27B, released on November 6, 2026. Nous fine-tuned it “to reinforce Western values on a Chinese-trained model,” and published the weights on Hugging Face under Apache 2.0, along with the training data and the evaluation set.
How much did Arete actually change Qwen’s political answers?
According to Nous’s self-reported evaluation, answers adopting the Chinese state’s account fell from 22% in the base Qwen3.8-27B to 0.9% in Arete with thinking off (0.6% with thinking on), while the average fact score rose from 2.32 to 2.98 out of 5. The biggest shift was in freedom of expression, where the fact score rose from 0.89 to 2.44 and the propaganda rate fell from 60% to 1%. These numbers have not yet been independently verified.
How did Nous retrain the model?
With supervised fine-tuning only — about 52.8 million tokens, 2 epochs, 1,202 steps, and roughly 20,000 prompts, using Thinking Machines’ Inkling as the teacher model. Nous states there was “No RL, no preference optimisation, and no architectural modification.” The vision encoder was frozen and the training data was text-only.
Did the fine-tune hurt Arete’s general abilities?
By Nous’s benchmarks, no. Terminal-Bench 2.1 rose from 67% to 70%, and Hermes Bench moved from 65.77% to 65.60%, which is essentially unchanged. The model is the same engine with different political reflexes, not a smarter or dumber model overall.
Can I try Arete-27B myself?
Yes. The weights are free on Hugging Face under Apache 2.0, and the 803-prompt Propaganda_Eval — including the judge prompt and scoring scripts — is public, so you can rerun the evaluation yourself. Budget about 54 GB for the BF16 weights and a GPU setup that can load a 27B model. Run your own prompts too, not just the published eval, before you use it in a client-facing tool.
Is Arete-27B neutral now?
No, and Nous does not claim it is. Arete was deliberately trained to reinforce Western values, which is itself a viewpoint. Its fact score of 2.98 out of 5 still trails Mistral Large 4, Nemotron 3 Ultra, and GLM 5.3, and 216 of 803 answers still scored zero on factuality. Treat Arete as a steered model with public receipts, not as a neutral one, and test it on your own use case before you ship it.