Chinese Z.ai's GLM-5.2 tops open AI leaderboards after U.S. export controls disable Anthropic models

At a glance:

• Z.ai's GLM-5.2 open-weight model tops public AI leaderboards within a week of release
• U.S. export controls force Anthropic to disable Fable 5 and Mythos 5 worldwide
• GLM-5.2 trained entirely on Huawei Ascend chips with no Nvidia hardware

U.S. export controls trigger immediate Anthropic shutdown

On June 12, the U.S. Commerce Department issued an export-control directive that barred Anthropic from supplying its Fable 5 or Mythos 5 models to any foreign national. The order effectively forced the San Francisco-based company to disable both models worldwide, cutting off access for developers and enterprises outside the United States almost overnight. The move marked an unprecedented escalation in the use of export regulations to restrict the global distribution of advanced AI systems, and it left a vacuum in the market for high-capability models that could be legally accessed by international users.

The directive did not target a specific country but applied a blanket restriction on foreign nationals, meaning even allies and partners lost access to Anthropic's flagship offerings. Industry observers noted that the decision appeared aimed at preventing model weights or capabilities from reaching Chinese entities, yet the collateral impact was global. Anthropic has not publicly detailed a timeline for restoring access, and the company's compliance posture suggests the shutdown could persist until a licensing framework is established — a process that typically takes months.

Z.ai launches GLM-5.2 on Huawei silicon the very next day

Beijing-based Z.ai, formerly known as Zhipu AI, began rolling out GLM-5.2 on June 13, just one day after the Anthropic restrictions took effect. The company released the model under a permissive MIT license, making its weights freely downloadable and usable for commercial and research purposes without the gating or usage monitoring that characterizes many Western open-weight releases. Z.ai stated that GLM-5.2 was trained entirely on Huawei Ascend chips, with no Nvidia hardware involved in the training pipeline — a claim that, if verified, would represent a significant milestone for China's domestic AI compute stack.

The model's architecture and training details have not been fully disclosed, but early benchmarks shared by Z.ai and independent testers indicate strong performance across reasoning, coding, and multilingual tasks. The use of Huawei Ascend processors — specifically the Ascend 910B series, according to supply-chain sources — underscores a growing effort by Chinese labs to decouple from U.S. semiconductor supply chains. Huawei's chip yield and software maturity have historically lagged Nvidia's CUDA ecosystem, but recent improvements in the CANN software stack and cluster-scale deployments appear to have closed enough of the gap for large-scale pretraining.

GLM-5.2 surges to top of open leaderboards as market value soars

Within a week of its release, GLM-5.2 climbed to the top of major openly available leaderboards, including the Open LLM Leaderboard on Hugging Face and the LMSYS Chatbot Arena's open-weight division. The model outperformed other openly licensed contenders such as Meta's Llama 3.1 405B, Alibaba's Qwen 2.5 series, and previous Z.ai releases across a range of evaluation suites. The rapid ascent was fueled by widespread downloads — Hugging Face reported over 200,000 pulls in the first five days — and enthusiastic community fine-tuning efforts that produced specialized variants for coding, math, and Chinese-language tasks within days.

The market reaction was equally dramatic. Z.ai's valuation, tracked through its Hong Kong-listed affiliate and private funding rounds, surpassed HK$1 trillion (approximately US$128 billion), making it the most valuable AI startup in China and one of the highest-valued private AI companies globally. Investors cited the combination of technical credibility, open distribution, and geopolitical tailwinds as drivers. The surge also lifted shares of Huawei's supply-chain partners and renewed debate in Western capitals about the effectiveness of chip export controls when domestic alternatives can support frontier-scale training.

Geopolitical implications reshape the global AI access map

The convergence of the Anthropic shutdown and the GLM-5.2 release has created a striking dynamic: for many users outside the United States, the most capable model they can legally download and run locally now comes from a company on the U.S. Entity List. Z.ai was added to the list in 2024 over alleged ties to the Chinese military, a designation that restricts U.S. persons from engaging with the firm without a license. Yet the MIT license on GLM-5.2 places no legal barrier on non-U.S. persons downloading, modifying, or deploying the model — a loophole that export-control experts say highlights the difficulty of regulating open-weight artifacts.

Policymakers in Washington, Brussels, and Tokyo are now grappling with whether future controls should target model distribution platforms, cloud inference services, or even the open-source licenses themselves. The European Union's AI Office has signaled it will assess whether GLM-5.2's availability triggers systemic-risk obligations under the AI Act, while Japan's METI has convened a working group on "open-model supply-chain resilience." For enterprise buyers, the episode accelerates a shift toward multi-model strategies that reduce dependence on any single jurisdiction's regulatory whims.

What to watch next: licensing, enforcement, and the next training run

The immediate focus shifts to whether Anthropic can secure a license to re-enable Fable 5 and Mythos 5 for foreign users, and on what terms. Commerce Department officials have hinted at a "validated end-user" framework similar to that used for semiconductor equipment, but no timeline has been published. Meanwhile, Z.ai's roadmap includes a 1-trillion-parameter dense model slated for late 2025, which the company says will also target Huawei Ascend clusters. If that training run succeeds, it would further erode the narrative that U.S. chip controls can durably constrain Chinese frontier-model development.

Researchers are also scrutinizing GLM-5.2's safety alignment and potential dual-use capabilities. Early red-teaming by independent labs has surfaced some jailbreak vectors, though none deemed critical. The MIT license means Z.ai has no contractual leverage to enforce responsible-use policies downstream, placing the onus on deployers and platform hosts. As the open-weight frontier moves east, the global AI governance conversation will increasingly center on whether transparency and community oversight can substitute for the centralized guardrails that characterized the first wave of generative AI deployment.