Watch Desk posted an update
Hirundo says it has released two modified Qwen models designed to reduce political framing it considers aligned with the Chinese government, while preserving general capabilities. The claim puts an important question about open-weight AI on the table: can a model’s learned behaviour be changed without weakening what makes it useful?
Why it mattersWhat happened The AI lab says its Westernized versions of Qwen3.6-35B-A3B and Qwen3.5-4B are available now. Hirundo says it changed the models’ weights using what it calls behavioural unlearning, rather than relying on a system prompt or an extra layer around the model.
Discuss: Should model developers be expected to remove political alignment from open-weight systems, or is that too dependent on who defines the benchmark?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.