Nous Research Strips Alibaba's Qwen of State Censorship in Arete-27B

Alibaba's open-weight models have long offered top-tier performance on a budget, but their baked-in state censorship has made them a risky bet for production software. US startup Nous Research claims to have solved that dilemma by releasing Arete-27B, a fine-tuned version of Qwen3.8-27B that slashes political bias from 22 percent to under one percent without degrading base performance, according to The Decoder.

What happened

The Decoder reports that Nous Research took Alibaba’s Qwen3.8-27B and ran supervised fine-tuning across roughly 53 million tokens over two epochs, deliberately skipping reinforcement learning and preference optimization. To evaluate the results, Nous created a benchmark of 803 prompts spanning seven sensitive topics, including free speech, surveillance, and democracy, using Elon Musk's Grok 4.7 as an automated judge.

According to Nous, unmodified Qwen3.8-27B scored 2.32 out of 5 on factual accuracy, with 22 percent of its outputs flagged as reproducing Beijing's official state narratives. In contrast, Western models like Mistral Large 4 and Gemma 4 31B registered propaganda rates below one percent, while other Chinese alternatives like DeepSeek V4.1 Flash and GLM-5.3 hit 16 and 15 percent, respectively.

To cleanse the model, Nous selected Inkling—an open model from Thinking Machines Lab—as its teacher model. The resulting checkpoint, dubbed Arete-27B, pushed factual accuracy up to 2.98 while driving the propaganda rate down to 0.9 percent. In the free speech category specifically, the factual score jumped from 0.89 to 2.44 while propaganda plummeted from 60 percent to just one percent. Crucially, raw utility barely budged: on the team's internal Hermes Bench, Arete-27B scored 65.60 percent compared to 65.77 percent for the stock base model. The weights and datasets have been released under an Apache 2.0 license on Hugging Face.

Why it matters for developers building with open models

If you run lean operations, open-weight Chinese models like Qwen and DeepSeek have been an open secret: they punch wildly above their weight class and cost almost nothing to self-host compared to commercial API giants. But using them in client-facing applications has felt like walking a tightrope. Beyond embarrassing public hallucinations on geopolitical prompts, The Decoder highlights research from CrowdStrike showing that DeepSeek generated less secure code when queries touched on sensitive Chinese political topics.

Arete-27B proves that this bias is not baked into the foundational architecture of the model—it is, as Nous puts it, a shallow layer applied late in training that can be stripped out with modest compute. For small teams that cannot afford proprietary API price tags but cannot risk rogue refusals or compromised security logic in their workflows, this model shifts Qwen from an awkward internal prototype to a viable production candidate. You get Alibaba's raw engineering muscle without Beijing's editorial guidelines, packaged under a commercial-friendly Apache 2.0 license.

What to do this week

1. Check your Hugging Face or local inference pipeline to evaluate whether Arete-27B can drop into existing Qwen-based deployments as a direct replacement.

1. Run a spot-check of your current open-source model prompts against edge cases in governance, compliance, or privacy to see if your system triggers unexpected refusals or sanitized answers.

1. Review the newly released Nous Research dataset on Hugging Face if you maintain proprietary fine-tunes, using their supervision approach to patch political bias out of your own open-weight pipelines.

Sources

• State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away — The Decoder