US officials are weighing restrictions on Chinese open-weight AI models after Moonshot's 2.8 trillion-parameter Kimi K3 outperformed leading American systems, drawing sharp pushback from researchers.
Key Points:
- Kimi K3 topped US frontier models on a major coding benchmark, reviving White House interest in curbing Chinese open-weight releases.
- The Commerce Department has not moved forward, and Moonshot plans to publish the full weights within days.
- Researchers argue a ban would shield American labs from competition without making AI any safer.
Kimi K3 Benchmark Revives Chinese AI Ban Push
Moonshot released Kimi K3 in mid-July, and the model surpassed OpenAI's GPT-5.6 Sol and Anthropic's Claude Fable 5 on the FrontierSWE coding benchmark with a score of 81.2. Axios reported on Jul. 20 that parts of the Trump administration are pressing again for de facto bans on foreign open-source models, citing cybersecurity risks. Sources close to the administration said Commerce weighed adding Chinese AI labs to its Entity List last year, a step that would cut off American access without a license.
Dean W. Ball, head of strategic futures at OpenAI, argued that Washington should manufacture regulatory uncertainty around the new models, since cheap open weights discourage capital spending at frontier labs. He walked that claim back within days after Yann LeCun, Martin Casado and other researchers said open software speeds up progress and can coexist with proprietary work.
A separate report from Politico said Commerce would not act soon.
Also Read: Dow Jones Industrial Average Volatility Tests Bitcoin's Decoupling Story
Researchers Question The Security Case
Braden Hancock, co-founder of Snorkel AI, said frontier-caliber open models will squeeze margins and pull down prices at the closed labs, which is why American companies keep adopting them. He does not expect total AI usage to fall as a result.
Sam Bresnick, a research fellow at Georgetown's Center for Security and Emerging Technology, questioned why federal power should shield American firms from rivals locked out over their origins.
Clem Delangue, chief executive of Hugging Face, said restrictions would hide risks instead of reducing them and would concentrate control among a few firms. Hancock added that US graduate programs now build mainly on open Chinese weights, shifting research gravity toward Beijing.
Chip Controls Offer A Narrower Lever
Any restriction faces a practical problem, because Moonshot plans to publish the full model weights on Jul. 27. Once those files circulate across mirrors and download sites, anyone with servers can run the system without touching an American cloud provider or paying a US licensing fee.
Bresnick said halting sales of Nvidia H200 processors to China would slow Beijing more directly than any software ban, and would avoid a fight over tools many US firms already run. Washington has spent four years tightening the hardware layer rather than the model layer, starting with sweeping semiconductor export controls in 2022.
Commerce also barred its own staff from using DeepSeek on government devices in March 2025, a device-level measure that never extended to the wider market.
Read Next: Meteorite That Hit A Home In New Jersey Fuels AI-Driven Hunt For Life