Z.ai releases GLM-5.3 weights with cyber-review licence for big firms

What happened: Z.ai (Zhipu AI) released the open weights of its GLM-5.3 model on Hugging Face on 28 August, two weeks after launching it by API. The delay let the Chinese lab finish a safety evaluation of the model's cyber capabilities.

The numbers: GLM-5.3 scores 84.5% on CyberGym, up from 77.2% for GLM-5.2, and ahead of Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). On ExploitBench it jumps to 54.4% from 24.4%. In testing, Z.ai says the model produced 2,436 vulnerability findings across 269 open-source projects, and completed 130 exploitation tasks in six hours versus 39 for its predecessor.

The details: The flagship model is a mixture-of-experts system with around 753 billion parameters, available in BF16 and FP8, and runs on vLLM, SGLang, KTransformers and Transformers. A smaller variant, GLM-5.3-Flash (320 billion total, 18 billion active), shipped on 26 August under a plain MIT licence. Z.ai says the cyber gains came from scaled post-training, not a retrained base model.

The catch: The flagship no longer uses the permissive MIT terms of GLM-5 through 5.2. Its new GLM-5.3 Licence requires any company with more than $10 billion in revenue over a 12-month period to pass a Z.ai security review before hosting the model themselves. Individuals and smaller firms face no restriction; routing to the model via API is also unaffected.

The rivals: Moonshot and DeepSeek, two other major Chinese open-weights labs, keep more permissive licensing terms, according to The New Stack.

Why it matters: A Chinese open-weights lab is now staging releases and gating access on safety grounds, the same debate US labs have been having over increasingly capable cyber models.

Who's affected: Large hyperscalers wanting to self-host GLM-5.3 face the new review; everyone else, from individual developers to smaller firms, can download and run it freely.

What's next: Critics are questioning whether the licence is genuinely about safety or about controlling commercial use by big players, a tension unresolved as rivals stick with looser terms, per VentureBeat.