Nous Research Says Fine-Tuning Removed Most Political Propaganda from Alibaba’s Qwen Model

Nous Research has released Arete-27B, a fine-tuned version of Alibaba’s Qwen3.8-27B that the company says sharply reduces pro-Beijing political responses while preserving the base model’s general performance. In Nous’s internal evaluation, the share of answers classified as propaganda fell from 22% to 0.9% after training on about 53 million tokens.

The result matters because it challenges a common assumption about political behavior in Chinese open-weight models: that government-aligned responses are deeply embedded throughout the model. Nous argues instead that at least some of the behavior can be altered with targeted supervised fine-tuning, without retraining the model from scratch or applying reinforcement learning.

The findings remain primarily vendor-reported. The evaluation was designed and run by Nous, and the comparison does not establish that every Chinese model contains a similarly removable layer of political alignment.

What Nous changed in Qwen3.8-27B

Nous Research trained Arete-27B from Alibaba’s Qwen3.8-27B using approximately 53 million tokens over two epochs. The company says it used supervised fine-tuning only, with no reinforcement learning or preference optimization. It also reports removing overlaps between the training material and its test prompts.

The training focused on responses involving free speech, state ideology, education, privacy and surveillance, democracy, and the environment. Nous says the objective was to steer the model toward factual answers rather than responses that repeat, endorse, or normalize official Chinese government narratives.

The resulting model and related materials, including the test dataset and training data, have been made available on Hugging Face under an Apache 2.0 license, according to The Decoder’s account. That could allow researchers and model developers to inspect the approach, reproduce parts of the evaluation, or test whether the same method works on other open-weight models.

Nous describes the political behavior it observed in Qwen3.8-27B as a “thin” layer applied late in training. That conclusion is an interpretation from the company’s results, not an independently established property of Qwen’s training process.

The evaluation showed a large political shift

Nous tested both models with 803 prompts across seven politically sensitive categories. For each prompt, the company identified two to four factual points from dated source material, such as laws, news reports, or academic research. A separate judge model then assessed whether the answer included, omitted, or contradicted those facts.

The judge was Grok 4.7 from xAI. Nous says the judge did not know which model produced each response and was instructed to cite the wording supporting its decision. The company classified an answer as propaganda when it repeated an official Beijing narrative instead of presenting the relevant facts, endorsed that narrative, or treated it as equally valid despite contrary evidence.

On this test, Qwen3.8-27B scored 2.32 out of 5 for factual accuracy and produced propaganda in 22% of responses. Arete-27B reached 2.98 for factual accuracy, while its propaganda rate fell to 0.9%. The largest change came in the free-speech category: Nous reports that the fact score rose from 0.89 to 2.44 and the propaganda rate dropped from 60% to 1%.

The reasoning-mode result is less complete. Arete reportedly reached a factual-accuracy score of 3.07 with reasoning enabled, but Nous based that figure on only two of five planned runs. That makes it an early signal rather than a stable benchmark.

For comparison, Nous says Mistral Large 4 and Gemma 4 31B each produced propaganda rates below 1% on the same evaluation, while GLM-5.3 and DeepSeek V4.1 Flash scored 15% and 16%, respectively. Those cross-model comparisons should also be treated cautiously because they rely on a single company’s prompt set, definitions, and judge-model methodology.

General capability claims are also internal

Nous says removing the political behavior did not materially weaken general performance. On the company’s Hermes Bench, Arete-27B scored 65.60%, compared with 65.77% for the original model. On Terminal-Bench 2.1, the score increased from 67% to 70%, although Nous cautions that the test setups differed and does not present that result as a meaningful gain.

Those numbers suggest that targeted alignment changes do not necessarily require a broad loss of coding or general-purpose ability. They do not, however, amount to an independent safety or capability evaluation. The benchmarks were reported by Nous, and the available evidence does not show how Arete performs across a wider range of languages, domains, adversarial prompts, or production workloads.

The choice of teacher model also illustrates a practical constraint for model developers. Nous selected Inkling from Thinking Machines Lab because it says using an open model would have created concerns around terms of service restricting distillation. Nvidia’s Nemotron 3 Ultra was reportedly considered but not selected after Nous observed what it described as irregularities on China-related topics.

That process highlights a broader issue: synthetic-data pipelines can transfer political behavior between models. Nous says it has seen signs of similar contamination in its own projects, while Inkling itself used models including the Chinese Kimi K2.5 in its synthetic training data, according to The Decoder.

Why the result matters for model builders

For AI builders, the immediate lesson is methodological rather than political. A relatively small supervised fine-tuning run—53 million tokens against a model of this class—may be enough to change highly visible refusal and framing patterns without rebuilding the underlying model. That could lower the cost and time required to adapt open-weight systems for different markets or organizational policies.

The risk is that a lower propaganda score does not automatically mean a model is neutral, truthful, or safe. A model can stop repeating one government’s narrative while still displaying other ideological preferences, uneven factuality, or new failure modes introduced by the fine-tuning data. The use of a judge model with its own political characteristics also makes evaluation design important.

For enterprise buyers, the release offers a possible route for testing and adapting open models before deployment. But buyers would need to examine training-data provenance, licensing, multilingual behavior, privacy handling, refusal policies, and performance on their own sensitive use cases. A low score on Nous’s political benchmark is not evidence that Arete-27B is ready for regulated or high-impact applications.

The finding also adds pressure to model vendors to document how political alignment is created and measured. Previous studies cited by The Decoder, including work from NewsGuard, CrowdStrike, Aleph Alpha, and ValsAI, have reported uneven behavior among Chinese models on topics such as Taiwan, Tibet, Falun Gong, human rights, and internment camps. Those studies likewise suggest that results can vary by model, language, topic, and response format.

What to watch next

The most important follow-up will be independent replication of the Arete-27B evaluation. Researchers should test the released weights and data with judges from different providers, human reviewers, multilingual prompts, and new political topics not present in the training set.

It will also be important to see whether the method transfers to other Chinese open-weight models, including DeepSeek, GLM, and Kimi systems, without damaging reasoning, coding, or factual performance. Longer-term tests should examine whether the altered behavior persists after extended conversations, tool use, system-prompt changes, and adversarial prompting.

Finally, model users should watch for evidence from real deployments rather than relying only on benchmark scores. Nous’s planned enterprise push around Hermes for Businesses may provide more information about how the company measures reliability, governance, and customization outside its research evaluation.

Creati.ai perspective

Arete-27B is a useful demonstration that political behavior in at least one tested model can be modified more surgically than many buyers might expect. The reported capability retention is notable, but the political conclusions are stronger than the evidence currently supports: one model, one internally designed test, and mostly vendor-reported results do not prove that ideological alignment is generally shallow or easy to remove.

For builders, the practical takeaway is to treat fine-tuning as an intervention that requires measurement, not as a guaranteed neutrality switch. The value of Arete-27B will depend less on its headline propaganda reduction than on whether independent evaluations confirm that the change is durable, broadly factual, and free of costly regressions elsewhere.