Nous Research has released the open-source Arete-27B — a fine-tuned version of the Chinese Qwen3.8-27B from Alibaba, which underwent cheap supervised fine-tuning on politically sensitive topics. After this, the share of answers containing official CCP propaganda dropped from 22% to 0.9%, while general agentic and coding capabilities remained at the base level. The weights are already available on Hugging Face and serve as a drop-in replacement for the original model with the same 27B parameters and 64k context.

image
image
image

What happened

The Nous Research team published the announcement "Introducing Arete-27B: Qwen3.8-27B Post-Trained To Reinforce Western Values" and opened the model weights. The recipe is supervised fine-tuning on approximately 52.8 million tokens: about 20,000 generated questions across seven topics (freedom of speech, state ideology, state and free market, education, surveillance and privacy, democracy, ecology), without reinforcement learning and without architecture changes. The donor teacher for distillation was the open-source Inkling model from Thinking Machines — it was chosen specifically to avoid violating the terms of use of closed APIs. Behavior was measured on the Propaganda_Eval set: 803 questions, each accompanied by 2–4 verifiable facts from dated sources, with LLM Grok 4.7 acting as the judge for the answers. The factual score increased from 2.32 to 2.98 out of 5, and the share of answers with the party line dropped from 22% to 0.9%. The freedom of speech topic improved the most: from 0.89 to 2.44, with propaganda dropping from 60% to 1%. General capabilities barely shifted: on Hermes Bench, the model showed 65.60% versus 65.77% for the base model — a difference within the noise range, while Terminal-Bench-2.1 even increased from 67% to 70%.

Context

Qwen3.8-27B is an open-source model from Alibaba, widely used as a base for local deployments, and, like other Chinese open-source models, carries a value layer tuned to the official positions of the PRC: refusals, evasions, and party-line formulations on sensitive questions. The Nous experiment showed the main point regarding the depth of alignment: this layer turned out to be superficial and can be removed by cheap supervised fine-tuning, barely affecting the capability layer. Methodologically, this is reverse value post-training — replacing the party line with declared Western values is symmetrical to the original alignment; it is a different configuration, not neutrality. A side result is the first comparative figures in the open field by the same judge: 15% of answers with propaganda markers for GLM 5.3 and 16% for DeepSeek V4.1 Flash, meaning the problem is not unique to Qwen. Nous itself emphasizes that the surface is broader than politics: it cites studies where the geopolitical context of a request worsened even code safety.

Why this matters for the industry

For teams building products on open Chinese weights, specific de-risking has been demonstrated: the value layer can be considered a separate, cheaply replaceable component, not an inherent property of the checkpoint. Arete-27B is a drop-in replacement for Qwen3.8-27B: parameters and architecture match, so the serving stack, quantizations, and pipelines do not need to be changed, and the recipe with an open-source teacher model is reproducible in-house without violating closed API terms. An operational nuance: each update to the upstream base resets the post-training, and value tuning will have to be repeated after each release. Further — these are expectations, not established facts: if the result is reproducible, a similar procedure will be applied to other entrenched behavior models, from event censorship to excessive refusals; the distribution of open weights will prospectively fragment into base checkpoints plus post-trainings for value profiles and jurisdictions — roughly like today's quantizations for different hardware — and eval sets on sensitive topics in the spirit of Propaganda_Eval may become part of standard CI alongside coding benchmarks.

Why this matters for users

If you are running Qwen3.8-27B locally and encounter refusals, evasive answers, or party-line formulations on politically sensitive questions, Arete-27B is a ready alternative: weights are downloaded from Hugging Face, deployment remains the same, just replace the checkpoint. It makes sense to run A/B on your own cases plus the open Propaganda_Eval to compare behavior before migration, rather than trusting only someone else's report. The practical effect for products is fewer refusals and evasions on topics of freedom of speech, surveillance, and democracy, with capabilities at the base level. One caveat: 0.9% is a one-time measurement obtained by the developers' judge, and is not a guarantee of quality on your data and your formulations.

What is currently unknown / limitations

Key figures were obtained by the developers themselves; independent verification is not yet available. Propaganda_Eval was compiled by the same team that conducted the experiment, and its questions relate to the same seven topics from which the training sample was generated; train/eval overlap was not checked in the sources, so final metrics may be inflated. The "propaganda share" metric is provided by the LLM judge Grok 4.7 without described human verification — the assessment is sensitive to the judge's prompt and its own value tuning, so inter-model comparisons are correct as relative, but not as absolute. The formula "drop-in replacement, costs close to zero, recipe replicable by any team" is exaggerated: among the explicitly noted unverified conditions are the licensing terms of the base model from Alibaba, on which Arete is built. Finally, the construction of "2–4 facts from dated sources" for each question is the design of the eval authors, not an independent metric of objectivity in controversial topics.

Sources

Author

Look at AI, editorial team