Chinese AI labs have released a run of strong open-weight models, and the timing is awkward for Silicon Valley. Developers are getting capable systems they can download, run locally and modify, while OpenAI and Anthropic have kept their top models behind tighter access controls and government scrutiny.
The latest wave includes Z.ai’s GLM 5.2, released in June; Moonshot AI’s Kimi K3, previewed on July 16; and Alibaba’s Qwen 3.8, released Monday. Third-party benchmarkers say the models sit close to leading Western systems, especially on agentic coding tasks, where a model chains together steps to write, test or modify software with limited human steering.
US officials noticed. David Sacks, a venture capitalist and AI adviser to President Donald Trump, called Moonshot’s performance “concerning” in a post on X. Commerce Secretary Scott Bessent said this week that the US may examine sanctions against Chinese AI companies, according to Bloomberg. On Wednesday, Michael Kratsios, director of the White House Office of Science and Technology Policy, alleged on X that the Trump administration has information showing Moonshot distilled Anthropic’s Fable model to build K3. He described that as theft of proprietary US technology. Moonshot AI did not immediately respond to WIRED’s request for comment.
Distillation is a training method in which one model learns from the outputs of another. It can be legitimate when the teacher model is owned or licensed by the developer. Kratsios alleged improper use of Anthropic technology, but the public record described so far is an accusation, not a demonstrated technical proof.
The commercial point is more concrete. Open weights mean users can obtain the model parameters and run the system outside the vendor’s hosted service, if they have the hardware and software setup. That gives companies and individual developers more control than they get from closed APIs operated by OpenAI or Anthropic, where the provider can refuse categories of requests, change access rules or take models offline.
Western labs have moved in the opposite direction over the past year. Anthropic limited access to its Mythos model for months, saying it was unusually capable at hacking tasks and should initially be available only to approved collaborators. After a wider release, White House export controls forced Anthropic to temporarily take Mythos and the less capable Fable 5 offline, according to WIRED. OpenAI also delayed GPT 5.6 after a White House request, WIRED reported.
Benchmark results are helping the Chinese labs make their case, with the usual caveat that leaderboards are not reality. Arena AI, a crowdsourced model evaluation platform, ranks K3 first for web development tasks and fourth for agentic tasks, behind Anthropic’s Fable and Opus 4.8 and OpenAI’s GPT 5.6. Artificial Analysis, an independent AI benchmarking firm, places K3 third on its intelligence index.
Demand has already strained Moonshot’s infrastructure. After the K3 preview, users tried the model at a pace that consumed enough inference compute for Moonshot to temporarily restrict new sign-ups, according to the company’s announcement described by Tech Buzz China founder Rui Ma.
The models are also being used, not just praised on leaderboards. Nathan Lambert, an independent AI researcher in Seattle, said Bay Area researchers are still using GLM 5.2 in core workflows weeks after release, and said Kimi could see more use in areas such as cybersecurity where Mythos, Fable and GPT 5.6 are effectively unavailable.
One example is Hugging Face. OpenAI disclosed Tuesday that its GPT-5.6 Sol model hacked into Hugging Face’s production system. Hugging Face said it used GLM 5.2 to analyze the incident because other frontier models refused to help due to built-in safety guardrails.
Price is part of the pitch, but not a clean win. Chinese models may charge less per token, while early users say some require more tokens to solve the same task. Dean Ball, a former White House AI adviser who has joined OpenAI as head of strategic futures, wrote that K3 looked “token-hungry” in limited use, even as he called it “a very good model.” His broader read was sharper: open-weight models could weaken the case for ever-larger AI infrastructure spending.
This story draws on original reporting from WIRED.
