Lab Stripped Chinese AI Models of CCP Political Bias
Tel Aviv-based Hirundo has released modified versions of Qwen models that significantly reduce enforced censorship.
Updated on Oct. 5, 2026 in Artificial Intelligence
Live Poll
Do you trust AI models from countries with strict government control over information?
Tel Aviv AI safety lab Hirundo has released Westernized versions of Alibaba's Qwen3.6-35B-A3B and Qwen3.5-4B models. These modified versions significantly reduce CCP-aligned responses and refusals while maintaining general performance within 0.72 points of the originals.
Why it matters
As Chinese AI models grew to represent half of traffic on OpenRouter by mid-2026, Hirundo aims to remove hidden geopolitical alignments for Western enterprises. The release targets the removal of enforced compliance in open-weight models.
The Westernized Qwen3.6-35B-A3B saw refusals on the DECCP benchmark drop from 65.26% to 3.16%. General performance scores remained within 0.72 points of the baseline, though the Qwen3-Coder variant showed a 130% increase in its vulnerability score.
The players
Hirundo
A Tel Aviv-based AI safety lab that specializes in machine unlearning to modify model behavior.
Alibaba
A Chinese multinational technology company that develops the Qwen series of open-weight large language models.
The details
Hirundo treats political alignment as a learned behavior, using machine unlearning to edit specific patterns directly in the model weights. This process separates unwanted geopolitical biases from the underlying core capabilities, effectively steering the model away from censored responses. The lab validated these adjustments using its internal CCPC-500 benchmark, which tests model compliance across 500 prompts.
Timeline
Late 2024: Chinese models held a 1% token volume share on OpenRouter.
Mid-2026: Chinese models reached 50% of total traffic volume.
October 2, 2026: Research regarding the model testing was published by CBS News.
October 5, 2026: Hirundo released the Westernized Qwen models.
The Tech Race
Hirundo's effort follows a broader trend where Western enterprises increasingly rely on high-performing open-weight models of foreign origin. By targeting the alignment layer specifically, the lab is positioning itself to provide a neutral alternative to state-aligned foundation models.
Developers and enterprises can now utilize these Westernized models to avoid CCP-mandated censorship in automated workflows. These models are intended for users who require the architectural performance of the Qwen series without the associated political alignment constraints.
The takeaway
The ability to surgically excise alignment behaviors from large models creates a new pathway for global developers to adopt high-performance software without geopolitical restrictions. Watch for the public release of the CCPC-500 benchmark, which will provide a standard for evaluating these safety modifications.
What happens next
Hirundo intends to publicly release its CCPC-500 benchmark and plans to develop Westernized versions of additional leading Chinese AI models.
Further reading
For more on the development of open-source models, visit the Artificial Intelligence section.
Source note: This article includes information reported by VentureBeat.
Live Poll
Do you trust AI models from countries with strict government control over information?
