A model any government can own
On July 27, 2026, Moonshot AI released the open weights of Kimi K3, a 2.8-trillion-parameter model, for free download — the largest open-weight model ever published (SCMP). It scores 57 on the Artificial Analysis Intelligence Index, ranking third globally, trailing only Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol (Artificial Analysis). Any government, company, or individual can now download it, retrain it on local languages and law, and run it entirely on domestic infrastructure — without paying Moonshot a license fee or renting a GPU cluster from Microsoft, Google, or Amazon (Rest of World).
That matters because sovereign AI has, until now, been sold as a hardware problem with a software rental attached. Nations spend billions building data centers on their own soil, then lease compute and models from US cloud providers for years at a time. Kimi K3 breaks the software half of that bundle. As Mohammed Soliman of the Middle East Institute told Rest of World, "an open, high-quality model like Kimi K3 does change the calculation for governments" weighing whether to keep renting from Washington-aligned vendors.
The constraint that didn't move
What Kimi K3 does not change is who controls the silicon needed to run a 2.8-trillion-parameter mixture-of-experts model at scale. That remains gated by the same export licensing regime the Bureau of Industry and Security has run since its December 2, 2024 rule imposed sweeping controls on advanced semiconductor equipment, software, and 140 additional entities aimed at "the PRC's capability to produce advanced-node semiconductors" for AI and weapons systems (BIS). On January 13, 2026, BIS eased — but did not eliminate — that regime: license applications for Nvidia H200 and AMD MI325X chips to China and Macau now get case-by-case review instead of a presumption of denial, contingent on US-based third-party testing, KYC checks, and proof the export won't reduce domestic chip supply (BIS). That is a real loosening, but it is still a licensing chokepoint, not open access — the software is free; the compute to run it well is not.
This is the actual digital-trade story: the US lost its ability to gate model capability the moment Moonshot open-sourced weights that rival its own labs' output. What Washington still controls is model deployment at scale — and that leverage is eroding on its own timeline as China builds around it, with Alibaba's Zhenwu M890 and Huawei's Ascend 910C already adapted to run Kimi K3 domestically (SCMP).
Washington is split on the wrong question
The policy fight in Washington has focused on whether to restrict American access to Chinese open models, not on tightening the chip rules that are the actual point of leverage. Nvidia's Jensen Huang, joined by Microsoft, Google, Meta, and OpenAI, publicly backed continued open access; 179 Silicon Valley startups wrote to the administration urging the same (Rest of World). On the other side, Anthropic's Dario Amodei has warned that Chinese models could hand Beijing military advantage or enable repression, and hardliners including AI czar Michael Kratsios and Treasury Secretary Scott Bessent have gone further — accusing Moonshot of distilling Anthropic's Fable 5 and of acquiring banned Nvidia chips, with Bessent threatening sanctions for "distillation attacks."
The security case deserves a fair hearing before it gets dismissed. A frontier-capable model trained in part on distilled outputs from a banned competitor, or run on smuggled export-controlled hardware, is a legitimate sanctions target — that's a chip-smuggling and IP question, and existing entity-list and enforcement tools already reach it. Concerns about undisclosed censorship or backdoors in widely redistributed weights are also fair, even though officials have not documented any publicly in Kimi K3 specifically.
Enforce the chokepoint that still exists
Where the hardliners overreach is in treating a download ban as the remedy. Weights already on Hugging Face cannot be un-published, and 179 startups plus four of the largest US tech companies have already said a US restriction would mainly cost American developers access while doing nothing to stop adoption abroad. A ban on using Kimi K3 doesn't shrink China's model; it just walls American builders off from a top-three global system while everyone else — including US allies weighing their own sovereign-AI bets — keeps using it freely.
The coherent, proportionate policy is the one BIS is already running: keep licensing chip exports case-by-case, keep tightening entity-list enforcement against actual smuggling and distillation abuse, and let open-weight software compete on its merits. That preserves the one lever — advanced compute — that Kimi K3's release just proved is still scarce, instead of spending political capital trying to ban something that's already free.