Hirundo’s Westernized Qwen Raises a Harder Question Than Censorship

The release points to a growing strategic choice for developers using open-weight AI. They can adopt a model whose behavior reflects the priorities of its original developer, or use a modified checkpoint that promises fewer political constraints. That choice is not only about censorship. It also concerns provenance, evaluation standards and whether a model can be changed selectively without losing capabilities that are harder to measure.

Hirundo said in its announcement that it used behavioral unlearning to reduce what it describes as geopolitical alignment in Qwen models. The company’s method targets responses that it classifies as censored, propaganda-aligned or politically biased, with the stated goal of changing the model’s behavior through its weights rather than relying only on a system prompt or an external filter.

That distinction matters. Prompt-level controls can be removed, ignored or bypassed. A change to the weights, if it works as claimed, would be closer to editing the model’s default instincts. For companies deploying open models, that could provide a more consistent alternative to maintaining layers of moderation code around an otherwise unchanged checkpoint.

A dramatic result, built on a narrow test

Hirundo reports that its original Qwen3.6-35B-A3B produced a flagged response on 89.8% of prompts in its 500-question CCPC-500 benchmark. The Westernized version reportedly reduced that figure to 2.8%. Hirundo says the benchmark measures censorship, propaganda-aligned framing and political bias.

Those numbers are striking because they present the intervention as highly targeted. If the model truly changes political behavior while preserving general competence, developers could gain a version better suited to international products, research and applications where users expect answers that are not shaped by a particular government’s preferred framing.

The company also reports that reasoning, coding and instruction-following performance remained essentially unchanged. Its evaluation includes the DECCP and ChinaBench tests, along with safety assessments, according to the released Hugging Face model card for Qwen3.6-35B-A3B-Westernized. The model card identifies the checkpoint as a merged, standalone version based on Qwen3.6-35B-A3B and reproduces Hirundo’s reported benchmark results.

For developers, the release is therefore more useful than a simple political statement. It offers a downloadable artifact that can be tested, integrated and compared with the original model. That creates the possibility of independent scrutiny, but it does not establish that the company’s conclusions are correct.

Neutrality is not a universal measurement

The central limitation is the benchmark itself. Hirundo selected the prompts, established the labels and decided which answers counted as politically aligned or censored. Those choices can strongly influence the outcome, particularly when the distinction between factual caution, political bias and neutral language is subjective.

Yash Thakker of explainx.ai argued in an audit of the release that the 89.8% and 2.8% figures depend on Hirundo’s own prompts and labeling. Thakker also questioned the breadth of the capability testing and pointed to Hirundo’s commercial incentive, calling for independent reproduction and broader evaluation.

That criticism is the only opposing assessment identified for this story. No independent source provided here has confirmed Hirundo’s headline numbers or demonstrated hidden capability loss in the modified model. That absence does not disprove the release, but it limits what can responsibly be concluded from it.

A model may perform normally on coding or instruction-following tests while changing in less visible ways. It could become less willing to discuss politically sensitive subjects, more likely to hedge, or less accurate when historical and political context overlaps. General benchmarks may not reveal those effects. Nor would a test focused on short answers necessarily capture changes in long-form reasoning, retrieval, multilingual behavior or unusual user prompts.

The market implication

The release reflects a larger shift in open AI. Model weights are increasingly becoming strategic raw material that different companies can adapt for distinct regulatory, cultural and commercial environments. One organization may optimize a model for political compliance, another for fewer refusals, and a third for safety or corporate policy.

That flexibility is valuable, but it moves a previously hidden decision into the open. Developers will need to know not only what a model can do, but who changed it, what was removed and how the changes were measured. A modified checkpoint can be more suitable for a target market while also making the original model’s provenance harder to evaluate.

Hirundo’s release is therefore best viewed as a testable claim, not a settled result. Its reported figures suggest that politically sensitive behavior may be altered without obvious damage to several general capabilities. Whether that achievement survives independent benchmarks, adversarial prompts and real-world deployment will determine its significance.

The future competition may not be between one model and another. It may be between competing versions of the same model, each presenting its own definition of acceptable neutrality.