Hirundo Releases Westernized Qwen, Cutting Politically Biased Answers from 89.8% to 2.8%

Hirundo, an AI safety lab based in Tel Aviv, has released Westernized versions of Alibaba's Qwen open-weight models, using machine unlearning to remove political alignment directly from the model weights. The models, Qwen3.6-35B-A3B-Westernized and Qwen3.5-4B-Westernized, are now available on Hugging Face.

In Hirundo's evaluation on a 500-prompt benchmark, the original Qwen3.6-35B-A3B produced CCP-aligned censorship, propaganda-framed responses, or political bias in 89.8% of answers. The Westernized version reduced that to 2.8%. External benchmarks showed refusals on DECCP dropped from 65.26% to 3.16%, and non-compliance on ChinaBench fell from 96.67% to 6.67%. General capabilities remained largely intact, with average scores on GPQA, IFBench, LiveCodeBench and MMLU-Pro staying within 0.72 points of the original model. Safety and harmfulness benchmark scores also held.

The company said the change is made in the weights rather than via a system prompt, so it persists in any downstream use. Hirundo also published a technical report and plans to release its CCPC-500 benchmark publicly. CBS News independently tested both models in its report broadcast on October 2.