Hirundo’s Westernized Qwen Tests Whether AI Values Can Be Rewritten

Imagine asking an AI assistant about a politically sensitive event and receiving an answer that is direct, historically grounded and free of rehearsed official framing. The same system then writes software, follows a complex instruction and explains a technical problem with the fluency expected from a modern open model.

That is the future Hirundo is presenting with its Westernized versions of Alibaba’s Qwen models. The project is also a test of whether artificial intelligence can be edited internally, rather than managed through visible prompts, keyword filters or a separate moderation layer.

Editing behavior inside the model

Hirundo’s announcement describes a method it calls weight editing and behavioral unlearning. Instead of adding a system prompt that tells Qwen how to respond, or placing an external filter in front of its answers, the company says it changed the model’s parameters directly.

Those parameters are the numerical structures learned during training. They shape how a model associates questions with concepts, language and likely answers. Editing them is therefore more ambitious than changing the instructions shown to the model at the beginning of a conversation. The goal is to alter a recurring behavioral pattern at its source.

Hirundo applied the approach to Qwen3.6-35B-A3B, an open-weight model, and evaluated the original and modified versions against 500 sensitive prompts. The company says its test measured censorship, propaganda and geopolitical framing associated with the Chinese Communist Party.

According to Hirundo, the original model produced a response classified as showing CCP-aligned censorship, propaganda or political bias in 89.8 percent of cases. In the Westernized version, that figure fell to 2.8 percent.

Those numbers describe Hirundo’s own evaluation, not a settled industry measurement. The result depends on how the prompts were selected, how responses were classified and what counts as political bias rather than ordinary caution, uncertainty or factual disagreement.

Capability preserved, according to reported tests

The central claim is not simply that the model gives different political answers. It is that the change did not meaningfully damage its general abilities.

Hirundo reports that reasoning, coding and instruction following remained essentially unchanged after the editing process. The company also compares the modified model with external benchmarks in its announcement, presenting the work as an attempt to separate geopolitical behavior from broader language model capability.

A model that loses politically aligned framing but also becomes less accurate, less coherent or less reliable would offer limited practical value. Developers would face a tradeoff between changing the model’s behavior and preserving the performance that made the original model useful.

The project’s importance therefore lies in the proposed separation. If the method works beyond this evaluation, developers could potentially remove narrow behaviors while keeping a model’s competence intact. That could make open models easier to adapt for different regions, institutions and cultural expectations.

A checkpoint for independent scrutiny

Hirundo has published a standalone merged checkpoint and model card for Qwen3.6-35B-A3B-Westernized. The repository includes the reported CCPC-500 results, along with DECCP, ChinaBench and safety evaluation results.

Publishing the model allows researchers and developers to test the claim in settings beyond Hirundo’s prompts. They can compare answers across languages, rephrase sensitive questions, examine refusals and look for cases in which the editing introduced unexpected changes. They can also test whether the model’s political behavior shifts consistently or only on the patterns represented in the evaluation set.

That scrutiny matters because “bias” is not a single technical property. A response may be censored, propagandistic, diplomatically cautious, factually wrong or simply framed from a different political perspective. A credible evaluation must distinguish among those possibilities and make its scoring rules visible.

The broader issue reaches beyond Qwen. As open-weight models spread, model editing could become a form of ideological product design. One organization might remove Chinese state-aligned framing. Another might alter content it considers culturally unacceptable. A third might claim to be removing bias while quietly imposing its own.

The promise is a more adaptable model that behaves differently without forgetting how to reason. The risk is that weight editing turns a model’s internal worldview into an opaque battleground. Hirundo’s release provides a useful experiment, but independent replication will determine whether it represents durable machine unlearning or a narrowly tuned political fine-tune.