A Qwen AI search tool vanished from the Federal Register

A Chinese AI model was quietly doing the searching on the Federal Register, and now it is quietly not. US government officials removed the Qwen-powered search option from the Federal Register website on Wednesday 16 September, according to Ars Technica, which cites Reuters. The tool let visitors search public comments on proposed regulations using an open-weight model from Alibaba's Qwen family.

What was actually running on the site

The evidence is patchy, because nobody at the National Archives has said a word about it. A widely shared screenshot posted on X on 15 September showed the search option sitting on the government website, and an archived copy of the site's source code confirmed the Qwen model was taken down a day later, so the service was live for at least 24 hours. Nobody has explained when it went up, who approved it, or why it was pulled. The National Archives, which runs the Federal Register, has not commented and did not respond to Ars Technica's questions. Neither have the White House or the FBI.

The timing could hardly be worse

Earlier in September the FBI named Alibaba among six Chinese firms it accuses of industrial-scale distillation, the practice of training new models on the outputs of American ones. The agency describes it as a way for the United States' biggest competitor to cut costs and development time in the race for AI leadership. Having a government records site quietly serve answers from an Alibaba model while federal law enforcement warns the public off such models is the kind of contradiction that writes its own headlines. Chinese outlet Sina.com reported that the Federal Register was running a small open-weight model at the Qwen3 0.6B level rather than a flagship, that it only retrieved documents instead of reasoning over them, and that no government data was sent to Alibaba or to third parties. If that is accurate, the security case for pulling it is thin.

Independent experts told Reuters the deployment was unlikely to have posed a substantial risk, for the straightforward reason that the material it searched is already public. The politics is where the heat sits. Daniel Castro, president of the Information Technology and Innovation Foundation, called the disconnect insane. Representative John Moolenaar, who chairs the House China Committee, said no federal government entity should use a Chinese AI model at all, arguing that doing so only deepens the government's dependence on them. Georgetown law professor Anupam Chander offered the counterpoint that the site's content is public, so the model was never working with sensitive information.

Why the size of the model is the whole argument

Strip out the noise and the real story is about scale. A 0.6B-parameter model is small enough to download and run on your own hardware, which is precisely why open-weight models from China have become a default choice for organisations that would rather not send their queries to somebody else's cloud. It is also why the Federal Register's brief experiment was, on the technical merits, a defensible way to search a public archive cheaply and privately. American export controls are built to slow frontier training by restricting access to the most advanced chips. They say nothing at all about a small retrieval model sitting on a government server, and that gap is exactly the one the Federal Register wandered into.

Our opinion

This is a rare episode in which almost everyone involved is right and the outcome is still unsatisfying. The FBI is right that wholesale dependence on Chinese models is a strategic problem. The National Archives is right that running a small retrieval model locally is a sensible, private and inexpensive way to search documents that are already public. Castro is right that the optics of the thing are absurd. What is missing is any policy that can tell the difference between a 0.6B search assistant and a frontier model trained on somebody else's outputs, and until Washington writes one, agencies will keep making these calls quietly and hoping nobody takes a screenshot. Pulling the tool solved the embarrassment rather than the question. A short, plain note on what was running, where the data went and why it was removed would have cost the National Archives nothing and told the public far more than the silence has.