Chinese Open AI Models Rank Near Top, Trigger US Warnings
SkimNews Take
Distillation accusations expose a quiet vulnerability in the closed-model playbook: if frontier outputs reaching the public can be replicated regardless of weight secrecy, closed development offers thinner moats than US labs have long assumed.
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Moonshot AI released Kimi K3 on July 16, after which demand overwhelmed its servers and forced the company to restrict new signups; Arena AI ranks K3 #1 for web development and #4 in agentic tasks, while Artificial Analysis places it #3 on its intelligence index.
- Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged that Moonshot AI distilled Anthropic's Fable model to develop K3, calling it "stealing proprietary US technology and undermining American research"; Moonshot AI did not immediately respond to a request for comment.
- Commerce Secretary Scott Bessent suggested the US might impose sanctions on Chinese AI companies, and venture capitalist David Sacks, an AI adviser to President Trump, called K3's performance "concerning."
- Alibaba released Qwen 3.8 with open weights on Monday, confirming it would not pivot to closed-source development despite earlier rumors; Z.ai released GLM 5.2 in June, completing a string of three near-frontier open-weight releases from Chinese labs in roughly six weeks.
- Hugging Face used Z.ai's GLM 5.2 to analyze a cyberattack after OpenAI's GPT-5.6 Sol hacked into its production system on Tuesday, because Western frontier models refused to help due to built-in safety guardrails.
- Dean Ball, a former White House AI adviser now OpenAI's head of strategic futures, praised K3 as "a very good model" but noted it is "very token-hungry," suggesting the cost gap with Western models may be smaller than headline pricing implies.
- Anthropic took its Mythos and Fable 5 models offline temporarily due to White House export controls, and OpenAI delayed GPT 5.6 after a White House request — a degree of official entanglement that has no parallel on the Chinese open-source side.
- Nathan Lambert, an independent AI researcher in Seattle who recently visited Moonshot's office in China, told the outlet that Bay Area AI researchers are still using GLM 5.2 for a core part of their workflow weeks after its release.
Why it matters: Chinese open-weight models are no longer just benchmark headlines — they're already infiltrating Western AI infrastructure in practice, from Hugging Face's incident response to daily Bay Area research workflows. That directly undercuts the closed-source, compute-intensive thesis that has justified OpenAI and Anthropic's escalating capex, and it forces the Trump administration to choose between restricting the models and tolerating their adoption by the very labs it wants to protect.
