The Pulse
Hirundo Releases Qwen Models With Reduced Political Censorship
Hirundo reports that weight edits sharply reduced political censorship-related responses in two Qwen models, with small changes on the capability tests it ran. Its safety results were mixed: some harmful-compliance and harmless over-refusal
AI.info Team · Written with AI ·
Hirundo says it has stripped politically aligned responses from two Alibaba Qwen models; its own results show a trade-off worth watching: the revised models sharply reduce targeted behavior, but some safety measures worsen. The company published the models and its evaluation on October 5, describing the work as “Westernizing” Qwen through changes to model weights rather than through an external filter.
Hirundo reports that its Qwen3.6-35B-A3B version reduced a censorship, bias and propaganda score from 89.8% to 2.8% on the company’s new CCPC-500 test. Results also fell on two other evaluations: refusal rates on DECCP went from 65.26% to 3.16%, while ChinaBench non-compliance went from 96.67% to 6.67%. Those figures are Hirundo’s reported results, not an independent audit of its method.
Weight edits target a learned behavior
Hirundo’s approach begins by prompting the base model on political topics and selecting responses that show the behavior the company wants to remove. It trains a LoRA adapter on those examples, uses a separate set of prompts to retain other behavior, then merges the adapter into the base model to create a standalone checkpoint.
“The hard part is not broadly removing the behavior (you can find abliterated models across the internet), but removing only the behavior without impact other parts of the model.”
Ilana Grinberg, AI Researcher at Hirundo, writes.
The larger release is Hirundo Westernized Qwen3.6-35B-A3B; the smaller is Qwen3.5-4B Westernized. Hirundo says both are available on Hugging Face. Because the adjustment is embedded in the weights, the company says, it carries into systems built on the checkpoint rather than depending on a system prompt or separate guardrail.
Capability scores move little; safety measures vary
On four general benchmarks for the 35B model, Hirundo reports that the largest decline was 1.83 percentage points on LiveCodeBench, from 56.39% to 54.56%. Scores fell by less than one point on GPQA, IFBench and MMLU-Pro. The results suggest limited changes on those tests, but do not establish that performance stays the same across other tasks or real-world uses.
Safety measurements were not uniformly better. The 35B model’s XSTest and OR-Bench safety results were measured with Thinking OFF. On OR-Bench, harmful compliance rose from 0.31% for base Qwen3.6 to 0.62% for Hirundo’s version—four responses versus two, according to the company—while XSTest harmful compliance declined from 18% to 17%. Harmless-prompt over-refusal shifted from 9.6% to 9.2% on XSTest and from 94.11% to 94.72% on OR-Bench.
The smaller model’s results also vary by setting. With reasoning turned on, its refusal rate on DECCP fell from 60.71% to 16.67%, a smaller reduction than when reasoning is off, when the rate fell from 57.89% to 6.32%. Across six capability evaluations with reasoning off, Hirundo reports an unweighted mean change of minus 0.36 percentage points. On the smaller model’s safety tests, XSTest harmful compliance rose from 20.50% to 22.00%, and harmless over-refusal rose from 5.20% to 6.40%. OR-Bench harmful compliance fell from 1.37% to 1.07%, while harmless over-refusal rose from 87.53% to 89.61%. These safety comparisons were made with Thinking OFF.
Hirundo’s new test is not public yet
The company’s headline result relies on CCPC-500, a 500-prompt benchmark it introduces in this release. Hirundo says the test covers topics including Taiwan, Tibet, Tiananmen, Xinjiang, Hong Kong and Falun Gong, and that it plans to publish the dataset and an Inspect AI benchmark soon. Until those materials are available, outside researchers cannot reproduce the company’s central comparison using that test.
Hirundo also compares its 35B model against Snowdon1.1-Small, an independent realignment of the same base model by Thomson Reuters and the FAIR Lab at Imperial College. On CCPC-500, Hirundo reports 2.8% for its model and 30% for Snowdon; on DECCP, 3.16% and 20.21%; on ChinaBench, 6.67% and 40%. The figures offer a comparison across approaches, though the benchmark’s unpublished status limits scrutiny of the first result.
The release puts a concrete claim in front of developers considering open-weight Qwen: political response patterns can be reduced through targeted weight changes, with relatively small shifts on the general capability tests Hirundo ran. The company’s published safety results also show why the claim needs to be assessed measure by measure. Its Westernized Qwen3.6-35B-A3B checkpoint is available on Hugging Face, while CCPC-500 is still forthcoming.