Start with the direct answer, because most coverage buries it: as of July 26, 2026, there is no US ban on using Chinese AI models. No law, no executive order, no restriction on downloading Qwen's weights or calling DeepSeek's API. What exists is something earlier in the pipeline and moving fast: a Treasury-led review of Chinese open-weight models for alleged IP theft, with sanctions explicitly on the table — and, within the same eight days, the largest pro-open-model pushback coalition American tech has assembled. This page tracks both sides with dates and sources, sketches what a restriction could realistically look like, and ends with the practical part: what to do about your stack. It is a policy tracker, not legal advice, and it updates as events land.
What Washington has actually said
The review is real and its language is unusually sharp. On July 21–22, Treasury Secretary Scott Bessent told reporters the administration is examining leading Chinese open-weight models for evidence they copied US frontier labs, saying: *"We are finding watermarks of our US large language models on many Chinese models, and that's unacceptable"* — and, of distillation: *"There's a very technical AI word for it called distillation, but you and I would call it theft."* The instrument he named is sanctions ("We have the ability to sanction them"), with findings expected "in the coming days or weeks."
The trigger, by every account, was Kimi K3's launch — the open-weight model that reshuffled the leaderboards and the chip market in the same week. Reporting around the review says White House officials allege Moonshot illicitly distilled Anthropic's Fable model and trained on restricted Nvidia GB300 hardware. Neither allegation has been publicly evidenced, and Moonshot has not been formally charged with anything.
What has not happened, precisely because the distinction matters: no finding has been published, no sanction imposed, no rule proposed that restricts private US use of these models today.
The pushback: eight days, one coalition
The counter-mobilization came faster and broader than the threat:
| Date | Who | What |
|---|---|---|
| Jul 21–22 | Treasury Sec. Bessent | Review announced; sanctions threatened |
| Jul 22 | ~200 startups (Little Tech Association — Y Combinator, Proton among them) | Letters to President Trump, Commerce Sec. Lutnick and OSTP Dir. Kratsios: do not ban or restrict access to Chinese open-weight models, naming Moonshot and Alibaba. Particle founder Suhail Doshi: hundreds of startups would *"instantly die"* if K3 access were blocked |
| Jul 22 | Nvidia CEO Jensen Huang (Axios) | Chinese open models are "excellent"; US firms should "absolutely" use them |
| Jul 24 | AMD CEO Lisa Su; 25 majors (Microsoft, Nvidia, IBM, Meta…) | Su backs open models; joint letter "Open Weights and American AI Leadership" |
| Jul 25 | Huang (first-ever X post), Pichai, Altman | Huang's debut post shares the Nvidia letter; Pichai: "very happy to support this on behalf of Google"; Altman signs, "open source and proprietary" |
The absence list is as telling as the signatures: Anthropic appears on none of the letters — a fact the community noticed loudly. It is also the alleged victim of the distillation in question, which makes its silence readable in more than one way; we note it without scoring it.
The economic logic driving the coalition is not sentiment. Small companies build on open weights because the price ladder makes frontier-adjacent capability affordable; cutting access reprices their entire cost base overnight. That is what "instantly die" means.
What a restriction could realistically look like
Labeled clearly: this section is analysis, not reporting. Restrictions come in very different sizes, and they land very differently on APIs versus weights:
- Sanctions on specific labs — the instrument Bessent actually named. Sanctions restrict *transactions with the sanctioned entity*: for US persons, paying a sanctioned lab's API would become the clear problem. Weights already published and downloaded are copies on other people's disks — a sanction does not reach into them retroactively, though it could block future releases' distribution channels.
- Government-scope bans — devices, agencies, procurement. Precedent already exists in multiple jurisdictions from earlier DeepSeek episodes. Real, but bounded: they do not touch private-sector use.
- Distribution pressure — app stores, US-based hosting platforms delisting models or apps. TikTok-style, messy, and the most visible-to-consumers scenario; least relevant to developers pulling weights or calling non-US endpoints.
- A true private-use ban — the scenario the 200-company letter is aimed at. It is also the hardest to construct and enforce: open weights are already mirrored globally, and the coalition now lobbying against it includes most of the US AI industry's own leadership.
The structural fact underneath all scenarios: openness is the hedge. A model whose weights you hold under an MIT license (GLM-5.2 today; K3 committed by July 27) cannot be un-downloaded by policy. APIs are where policy bites first; local weights are where it bites last, if ever.
What this means for your stack
Depends on where you sit:
- US companies: exposure is real but currently prospective. The rational moves are the boring ones — keep model access abstracted behind an OpenAI-compatible layer so any single vendor becoming radioactive is a config change; mirror the open weights your products depend on while licenses plainly allow it; watch the review's findings, not the rhetoric.
- Everyone outside the US — which is most of this site's readers: a US sanction or ban does not directly bind you. Second-order effects are possible (US-based clouds, app stores, payment rails), which is an argument for endpoints and suppliers that do not route through US chokepoints.
- Everyone, regardless: the events of this week are the strongest argument yet that the open-weight commitments matter beyond ideology. They are the difference between renting capability that policy can revoke and holding capability that it cannot.
The Chinese lineup — GLM-5.2, DeepSeek V4, Kimi, Qwen — runs today on Turiloop behind one OpenAI-compatible key, from outside US jurisdiction, with an international card. Routing around single points of failure is what the architecture is for.
FAQ
Is it illegal to use Chinese AI models in the US right now? No. As of July 26, 2026 there is no US law or order banning private use of Chinese models or their APIs. Restrictions that do exist are government-scope (official devices, some agencies). This page updates if that changes.
What exactly is the US government reviewing? A Treasury-led examination of leading Chinese open-weight models for evidence of IP theft — specifically distillation of US frontier models. Secretary Bessent says results are expected within "days or weeks" and that sanctions are the available instrument.
Would sanctions cut off API access to models like Kimi K3? For US persons, sanctions on a lab would make transacting with it — including paying for its API — the core legal problem. Already-published open weights sit differently: they are distributed copies, and policy cannot retroactively delete them. (Analysis, not legal advice.)
Does any of this affect developers outside the US? Not directly — US measures bind US persons and US-jurisdiction infrastructure. Indirect effects (app stores, US clouds, payments) are possible, which favors supply chains that do not depend on US chokepoints.
What is the safest way to build with Chinese models this year? Abstract the vendor: an OpenAI-compatible endpoint makes every model a config value. Prefer models with published weights and permissive licenses. And treat this page's timeline as the thing to watch — the review's published findings will matter more than any statement made before them.