AINEWS
Beijing's top labs are publishing their models' inner workings for anyone to download, even as Washington calls one release stolen American research and floats sanctions on the rest.
Three open-weight Chinese models landed in the second half of July 2026. Z.ai shipped GLM 5.2 in June, Moonshot AI followed with Kimi K3 in mid-July, and Alibaba released Qwen 3.8 on Monday, all available to download, inspect, and build on, in the same week that a senior White House official accused Moonshot of stealing American technology and a cabinet secretary floated sanctions on Chinese AI companies. US AI is heading the other way.
"Open weights" means publishing the model's trained parameters so anyone can run or fine-tune them locally, the difference between handing out a recipe and serving the finished dish. The Chinese releases cluster around a particular slice of the market: agentic coding, the use case where models write and run their own code to finish multi-step tasks, which has become the year's most-watched AI capability. Third-party benchmarks reported by Wired put the new Chinese models "nearly as well as the best Western frontier models" on those tests, a comparison the article attributes to outside evaluators without naming the benchmarks or methodology.
The US response has been twofold: legal threats and tighter doors. Michael Kratsios, director of the White House Office of Science and Technology Policy, said the Trump administration has "information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," calling it "stealing proprietary US technology and undermining American research." Distillation, in this context, means training a smaller model on the outputs of a larger one, a standard technique whose legality depends on the source lab's terms of use. Moonshot AI did not respond to a request for comment. Commerce Secretary Scott Bessent separately suggested the US might impose sanctions on Chinese AI companies, an extension of the export-control posture from chips to the company level. David Sacks, a venture capitalist and Trump AI adviser, called Moonshot K3's performance "concerning."
The Western contrast is a near-mirror image. Anthropic's Mythos model was described as so capable at hacking that only approved collaborators could use it, and the company's sister model, Fable 5, sat behind the same wall. White House broad export controls later forced both offline temporarily, a rare case of US policy determining which American frontier model an American company's customers could access. OpenAI delayed the release of GPT 5.6 after receiving a US government request, an unusual coordination between a frontier US lab and Washington over a launch.
The structural read: US AI is consolidating around closed, government-coordinated frontier capability, while Chinese AI is consolidating around open-weight distribution. The thesis is testable against the next round of releases, including any Moonshot or Alibaba response to the IP allegation, and against whether the US extends chip controls to model weights or distillation outputs.
The US closed posture has real justification: Anthropic's decision to wall off Mythos followed a documented safety and capability problem (the model was too good at hacking), and OpenAI's delay was tied to a specific government ask. "Closed" in this context is not secrecy for its own sake. It expresses a specific safety and capability stance. "Open" is doing more work as a label than the binary suggests. Meta's Llama lineage, the most prominent US open-weight project, ships under licenses that restrict commercial use at scale, and the new Chinese releases are open in the sense that anyone can download them, but trained on data whose licensing has not been independently audited.
This is the second open-weight shock in 18 months. The first was DeepSeek R1 in January 2025, which forced US labs to reckon with a Chinese lab matching frontier coding performance at a fraction of the compute. The second is broader. It is a cadence, not a release, with three Chinese labs shipping on a weekly clock. And it is a contest, not a capability gap, with a US administration publicly weighing sanctions, an IP allegation that names a specific technique, and a Western frontier contracting behind export controls and government requests at the same moment.
The next data point is the Kratsios allegation's resolution. A Moonshot denial, a US trade-remedy filing, or both, will set the rules for the next round of Chinese releases.