Nous Research Strips Political Bias From Alibaba's Qwen Model Using Targeted Fine-Tuning

To score outputs, the team utilized Grok 4.7 from xAI as an automated judge. Operating without knowledge of which model produced an answer, Grok assessed whether answers reflected, contradicted, or omitted the defined facts, flagging outputs as propaganda whenever official Beijing narratives were endorsed or treated as co-equal to verified facts. Under this benchmark, unmodified Qwen3.8-27B scored 2.32 out of 5 for factual accuracy, with 22 percent of its outputs containing propaganda. Western baselines like Gemma 4 31B and Mistral Large 4 scored under one percent, whereas Chinese models such as GLM-5.3 and DeepSeek V4.1 Flash registered propaganda rates of 15 and 16 percent, respectively.

To build the training data for Arete-27B, Nous selected Inkling from Thinking Machines Lab as its teacher model, opting for an open system to adhere to distillation terms. Nvidia's Nemotron 3 Ultra was considered, but Nous detected China-related bias in its outputs—an irregularity likely stemming from synthetic data derived from Chinese open weights (Thinking Machines itself incorporated Kimi K2.5 data into Inkling). The resulting training set of 53 million tokens was applied to Qwen3.8-27B across two epochs of supervised fine-tuning, excluding reinforcement learning and preference tuning.

Following the fine-tuning, Arete-27B raised its factual accuracy score to 2.98 (reaching 3.07 in reasoning mode across initial runs) and reduced its propaganda frequency to 0.9 percent. The most notable turnaround occurred in the free speech category, where the fact score moved from 0.89 to 2.44, and propaganda plunged from 60 percent to one percent. Overall capability was preserved: on the internal Hermes Bench, Arete-27B hit 65.60 percent versus 65.77 percent for the original Qwen base, while Terminal-Bench 2.1 recorded 70 percent compared to the base model's 67 percent.

Systemic bias across open-weight models

The persistence of political guardrails in Chinese open architectures has been tracked across several independent assessments. A 2025 investigation by NewsGuard found that five leading Chinese systems, including Qwen, amplified or failed to challenge pro-Beijing disinformation in 60 percent of evaluations. CrowdStrike observed that DeepSeek models produced less secure code when processing queries containing terms such as Falun Gong, Tibet, or Taiwan. Research by Aleph Alpha spanning 967 political queries found that Alibaba, DeepSeek, and Moonshot AI models offered balanced answers in only 17 to 41 percent of instances. A study from ValsAI highlighted differing stances within Chinese models, noting that GLM-5.3 Flash acknowledged human rights abuses, whereas Kimi K3 rejected reports of internment camps.

While Nous previously focused on neutrality with its 2024 Hermes 3 release—designing it to adapt dynamically to the system prompt's instructions—Arete-27B intentionally steers model outputs toward Western democratic values.

What it means for developers

For software engineers, Chinese open-weight architectures present an attractive combination of strong reasoning, high efficiency, and accessible weights, but their deployment often brings compliance and reliability hurdles when handling political, historical, or legal subjects. Nous Research's work demonstrates that these behavioral restrictions are not deeply integrated into the model's fundamental reasoning architecture. Instead, developers can potentially strip these guardrails using modest compute budgets and focused datasets.

Because Nous Research has published the weights, evaluation framework, and training data for Arete-27B under an Apache 2.0 license on Hugging Face, teams can inspect the supervision pipeline directly or adapt the recipe to their own pipelines. For engineering teams evaluating model options across ecosystems, developers can try top AI models cheaply through one API at https://apixoai.online, allowing them to benchmark alternative architectures such as Grok, DeepSeek, and Western foundation models under uniform conditions.

Backed by a recent $90 million Series B round led by Robot Ventures alongside Nvidia and Samsung at a $1.5 billion valuation, Nous plans to package these alignment methodologies into enterprise offerings through "Hermes for Businesses," establishing a commercial path for adapting international open-weight models to enterprise standards.

Source: State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away — The Decoder. Written by the Apixo team from that report.