After Washington blocked foreign access to Anthropic's flagship AI models Mythos 5 and Fable 5 over autonomous software exploitation concerns, MiniMax, Zhipu, and Moonshot released downloadable model weights for their flagship AI models M3, GLM 5.2, and K2.7 Code, opening…
On June 12, 2026, the U.S. government restricted access to Anthropic's newly released Mythos 5 and Fable 5 models to U.S. persons only, citing the risk that the systems could autonomously find and exploit software vulnerabilities. Within days, three Chinese AI labs, MiniMax, Zhipu AI, and Moonshot, pushed out their own flagship models, framing the releases as alternatives for the developers and researchers who had just been locked out of the U.S. frontier.
The Recode China AI newsletter, which tracked the rapid sequence, reports that MiniMax open-sourced M3, Zhipu released GLM-5.2, and Moonshot released K2.7-Code in the days following the suspension. All three labs explicitly positioned the releases as open-weight alternatives: downloadable files that let anyone run a capable model on their own hardware, with no API gatekeeper and no nationality check, according to Recode China AI.
"Open-weight" deserves a precise definition, because it is not the same as "open-source." When a lab releases open weights, it publishes the trained parameters of the model so researchers, startups, and safety teams can download them, run them locally, and inspect their behavior. The training code, the training data, and the safety fine-tuning pipeline usually stay proprietary. The distinction matters: open-weight models give users auditability and price leverage against closed APIs, but they also carry whatever biases, jailbreaks, or failure modes the lab did not remove before release.
The U.S. restriction itself was abrupt. Anthropic confirmed that it received the U.S. government's export control directive at 5:21 PM ET on Friday, June 12, 2026, and suspended all access to both systems for foreign nationals — including foreign-national Anthropic employees — immediately afterward. Mythos 5 and the safeguarded Fable 5 had launched only the prior week, and the newsletter describes them as Anthropic's most powerful models to date, rumored at roughly 10 trillion parameters — a figure Anthropic has not confirmed publicly. Anthropic cited cybersecurity concerns, specifically the ability of the models to autonomously identify and exploit software flaws, as the trigger for the federal action, and the company described the directive as an unexpected government intervention, as Recode China AI recounts.
MiniMax confirmed its open-weight release independently: at 4:01 AM ET on June 13, the company tweeted a link to the M3 model weights on HuggingFace, with the caption "M3 would never 🙂↔️ As a matter of fact, the weights are now open, too." The HuggingFace page lists M3 as a native multimodal model with approximately 428 billion total parameters, 23 billion activated per inference run, and a 1-million-token context window.
Zhipu's co-founder and chief scientist Tang Jie posted separately that GLM-5.2 is "fully open" and "frontier intelligence belongs to everyone," calling the U.S. restriction "deeply regrettable" and "non-technical." The Zhipu release announcement and Z.ai documentation confirm the model is available to GLM Coding Plan users, with a 1-million-token context window and a focus on long-horizon agentic tasks.
Moonshot released Kimi-K2.7-Code, a coding-focused model with 1 trillion total parameters and 32 billion activated per inference run, available on HuggingFace. The company claims substantial improvements over its predecessor on coding and agent performance benchmarks, and launched a Kimi Code Beta Program as a direct rival to Claude Code.
The structural shift underneath the news hook is what makes this moment worth attention. Capable AI that researchers, startups, and safety teams can actually run, inspect, and audit is now flowing more readily from Chinese open-weight releases than from the U.S. frontier labs whose access is narrowing for non-U.S. persons. The Recode China AI newsletter frames the three Chinese releases as a coordinated message: that their AI will remain open to the world while American frontier access contracts. That framing is one interpretation; whether the three labs coordinated in advance, or simply moved quickly in parallel after the U.S. announcement, is not established by the public record and should not be asserted as fact.
The framing has merit, and it has limits. On the merit side, a developer in Berlin, a red-team researcher in Bangalore, or a small lab in Nairobi can now download weights and run a frontier-class system locally, with no API contract, no monthly bill, and no export-control checkbox. On the limit side, the open-weight releases arrive without the safety guardrails, red-teaming, and misuse mitigations that closed U.S. labs argue are necessary in the first place, and the cybersecurity concern that triggered the U.S. move does not disappear when a similar capability is redistributed under an open license.
What to watch next is whether independent benchmarks confirm that M3, GLM-5.2, and K2.7-Code are genuinely competitive with the models they are positioned against, and whether the U.S. export-control posture is broadened to cover open-weight distributions, not just hosted API access. The Recode China AI story is the news hook. The unresolved question — whether a structural decoupling of the AI ecosystem, driven by U.S. policy, will outlast the news cycle — is the real story underneath it, as Recode China AI frames the moment.