Political risk / Jun 27, 2026 / 5 min
Chinese Labs Mailed Model Weights While Washington Gated Access
In the same June fortnight that Washington killed Fable 5, limited GPT-5.6 to roughly 20 vetted partners, and restored Mythos to only ~100 organizations, China's Zhipu, Moonshot, and Alibaba shipped GLM-5.2, Kimi K2.7, and Qwen 3.7 as downloadable open weights — proving the AI sovereignty race may be won by whoever mails the source code.
Washington is gating America's frontier models behind export controls and Commerce guest lists while Beijing ships rival weights to the world under MIT licenses — and developers outside the approved lists are already switching stacks.
The June split screen:
- June 12: Commerce ordered Anthropic to disable Fable 5 and Mythos 5 globally over a disputed jailbreak. Fable remains dark as of June 27.
- June 12: Moonshot AI open-sourced Kimi K2.7-Code — a 1-trillion-parameter coding model under a Modified MIT license.
- June 13–16: Zhipu AI shipped GLM-5.2 with a 1-million-token context window and full MIT open weights on Hugging Face.
- June 26: Commerce restored Mythos to roughly 100 approved organizations. OpenAI launched GPT-5.6 Sol to about 20 government-vetted partners.
The quotes that frame it:
- Zhipu founder Tang Jie launched GLM-5.2 the day after the Fable ban, calling the restrictions "deeply regrettable" and writing that "the path to AGI must never be enclosed by high walls."
- Lizzi Lee of the Asia Society Policy Institute told the South China Morning Post: "The market immediately understood the opening that Anthropic just created."
- CSIS analyst Kate Koren warned the controls give China "an opportunity to make inroads on international adoption" while U.S. frontier models average a seven-month lead.
What China shipped — and how close it is:
- GLM-5.2: Independent benchmarks rank it the top open-weights model on Artificial Analysis's Intelligence Index (score: 51). Zhipu reports 81.0 on Terminal-Bench 2.1 — within four points of Claude Opus 4.8's 85.0. Developer Simon Willison notes it ranks second on Code Arena's web-dev leaderboard behind only Fable 5.
- Kimi K2.7-Code: 1T-parameter MoE, 256K context, open weights on Hugging Face.
- Qwen 3.7 Plus: Alibaba's open-weight variant, downloadable for individual use.
The security paradox:
Anthropic warned in its June 12 statement that if a narrow jailbreak triggers a global recall, "it would essentially halt all new model deployments for all frontier model providers." Washington applied that standard to one American lab — while Chinese competitors shipped unrestricted weights the same week.
Meanwhile, Anthropic accused Alibaba of running 28.8 million distillation queries against Claude through fraudulent accounts — extracting capability through APIs while Commerce kept the models dark.
Convina's view: Export controls on hosted models do not stop diffusion — they stop trust. Every developer who watched Fable go dark and GLM-5.2 go MIT learned the same lesson: sovereignty in AI means owning the weights, not renting the API. Washington is securing America's frontier labs into a corner while Beijing builds the default stack for everyone Commerce leaves outside the guest list. Security policy that gifts market share to open-source rivals is not containment — it is a subsidy.