US Federal Register caught quietly testing Qwen-based search

The US Federal Register website was found to use technology based on Alibaba's open-source Qwen model in its search function. The finding came from a community user, was verified by independent researcher 96Stats, and was later covered by Reuters. Conclusion: the government was not "distilling" a Chinese model for generation; Qwen3 was used only for semantic search, on a tiny index that went live for about a day before being taken down—without the disclosure the site's own rules require. The timing—amid tightening US restrictions on Chinese AI, recent security-agency warnings about such models, and an FBI accusation that Qwen copied Anthropic—makes the contrast stark, and shows ordinary users can audit government AI usage with browser DevTools.

Confirmed

• On Sept 16, kimmonismus found the Federal Register search used a model distilled from Qwen, verified it personally with reproducible links, calling it deeply ironic.
• On Sept 17, 96Stats reconstructed the process via Chrome DevTools: the search form exposed internal search type IDs—3=semantic, 6=keyword, 306=Hybrid Qwen3, 914=Qwen3; the Qwen wording briefly vanished from the page but remained verifiable.
• Per 96Stats, Qwen was only used for semantic search—not for writing, answering, or touching regulatory text.
• 96Stats suspects the actual model is Alibaba's Qwen3-Embedding-0.6B, open-sourced in June 2025: strong for its size, self-hostable, no vendor dependency.
• The test index was tiny ("only a fraction filled") and was taken down after roughly a day.
• On Sept 18, Reuters reported the story, noting the FBI's prior accusation that Qwen copied Anthropic's technology.

Unconfirmed

• Whether the test violated the site's own disclosure rules: 96Stats says it was undisclosed; no official response yet.
• The exact model version remains 96Stats' inference, unconfirmed officially.
• The FBI accusation against Qwen has no official resolution; Reuters mentioned it only as background.

Why it matters

• The episode landed at a sensitive moment: US intelligence and cybersecurity agencies had just warned about such models, the Pentagon listed Alibaba as a military-linked company in June, and the FBI had accused Qwen of copying Anthropic—while Washington tightens curbs on Chinese AI, a US government system was actually using Chinese open-source model technology.
• Mainstream coverage by Reuters elevated a community-driven discovery into a public controversy.
• It also demonstrates that ordinary users can independently audit government websites' AI usage with browser developer tools, fueling the debate on public-sector AI transparency.

2026-09-16 ~ 2026-09-18 · 8 related posts

• [source] US Federal Register Search Found Using Distilled Qwen Models — kimmonismus · 2026-09-16
• US Federal Register's search mode found running distilled Qwen models — kimmonismus · 2026-09-16
• US government website found using Qwen3-based search despite intel warnings — 96Stats · 2026-09-17
• Federal Register site caught testing Qwen3-powered search, hidden from public view — 96Stats · 2026-09-17
• [source] Federal Register caught secretly testing Qwen3 search with a mostly empty index — 96Stats · 2026-09-17
• Pentagon listed Alibaba as military company in June — yet a federal site tests Qwen3 — 96Stats · 2026-09-17
• [source] US Government Website Used Chinese AI Search Tool Qwen That FBI Said Copied Anthropic — External_Mood4719 · 2026-09-18

1 near-duplicate retellings: 96Stats