# Z.ai ships GLM-5.3 open — but hyperscalers now need its permission

> Z.ai open-sourced GLM-5.3 under a licence gating hosting by firms over $10bn revenue.

*A Chinese lab's new licence forces companies over $10bn revenue through a security review before they can host its most capable open-weight model.*

By Bez · InsidersFeed
Canonical: https://insidersfeed.com/news/z-ai-ships-glm-5-3-open-but-hyperscalers-now-need-its-permission

Z.ai just gave the AI industry a new template for releasing dangerous capability as open weight: publish, but put a toll gate in front of the biggest players. The Chinese lab's GLM-5.3, launched by API on 14 August and released as open weights on 28 August, comes with a licence that requires any company with more than $10 billion in revenue over a 12-month period to pass Z.ai's security review before it can host the model itself. Smaller companies and individuals face no such restriction.

The gate exists because GLM-5.3 is unusually good at finding and exploiting software flaws. In Z.ai's own testing the model produced 2,436 findings across 269 open-source projects, prompting a two-week hold on the weights while the company ran a safety evaluation. On CyberGym, Z.ai's benchmark for offensive security tasks, GLM-5.3 scored 84.5%, ahead of both Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). On ExploitBench it jumped to 54.4% from 24.4% for the prior version, and in ExploitGym it completed 130 exploitation tasks in six hours versus 39 for GLM-5.2.

The flagship is a mixture-of-experts model at roughly 753 billion parameters, released in BF16 and FP8 and compatible with vLLM, SGLang, KTransformers and Transformers. A smaller companion, GLM-5.3-Flash (320B total, 18B active), shipped two days earlier under a plain MIT licence with no such restrictions. Z.ai says the cyber gains came entirely from scaled post-training, not a retrained base model.

The move is a departure for Z.ai, whose GLM-5 through 5.2 releases all carried permissive MIT terms. Rivals Moonshot and DeepSeek have kept their licensing loose, leaving Z.ai as the outlier among Chinese open-weights labs in restricting who can self-host its top model.

## The Insider view

The $10 billion threshold is aimed squarely at hyperscalers and the largest AI labs, the only parties with the infrastructure to self-host a 753-billion-parameter model at scale. That is a narrow but consequential set of customers, and Z.ai's review requirement gives it leverage over exactly the companies most likely to weaponise or redistribute the model's exploit-finding ability commercially.

Critics note the line between safety gatekeeping and competitive control is thin: a review clause that only binds the richest potential rivals also happens to slow down exactly the companies capable of building competing products on Z.ai's own research. Watch whether Moonshot or DeepSeek follow with similar terms, or whether they use continued MIT licensing to pull developers away from Z.ai's now-encumbered flagship. Watch too whether Western labs cite GLM-5.3's CyberGym lead as evidence that open-weight releases from Chinese labs are closing, or overtaking, the capability gap on offensive security tooling.

## Key takeaways

- Z.ai delayed GLM-5.3's weights two weeks over cyber-exploit capability before releasing on Hugging Face
- New licence forces >$10bn-revenue firms through a security review before hosting, unlike prior MIT releases
- GLM-5.3 beats Claude Mythos 5 and GPT-5.6 Sol on CyberGym, unsettling assumptions about Chinese labs lagging on capability

## Sources

- [Z.ai's GLM-5.3 goes open weight, but its new license aims at hyperscalers](https://thenewstack.io/zai-glm-weights-license/) — The New Stack, 2026-08-28
- [GLM-5.3 is here with advanced cyber capabilities](https://venturebeat.com/technology/glm-5-3-is-here-with-advanced-cyber-capabilities-and-reportedly-already-found-a-serious-vulnerability-in-cursor) — VentureBeat, 2026-08-14