Z.ai has published downloadable GLM-5.3 weights, but it did not carry forward the MIT licence used for GLM-5.2. The replacement retains broad permissions for developers and operators while adding a security-review condition aimed at the largest businesses that provide models as a service.

That distinction matters for companies deciding whether they can add GLM-5.3 to a commercial inference, fine-tuning or distribution stack. The condition is narrow enough that it will not apply to most developers, but organisations near its boundary should read the exact text rather than rely on an open-weights label.

What Z.ai released

Z.ai posted the GLM-5.3 model files and model card on Hugging Face on August 28, 2026. The repository lists a 753-billion-parameter model with downloadable weights, and its commit history records the initial release.

The model card says GLM-5.3 uses the same base model as GLM-5.2, with its claimed improvements coming from post-training. Its performance figures should be treated as vendor-reported results, not proof that it will outperform alternatives in a particular production environment.

The lasting operational change is the licence. GLM-5.2 is marked as MIT-licensed, while GLM-5.3 carries a new licence written specifically for this release.

The licence still grants broad rights

The GLM-5.3 Licence permits licensees to use, copy, modify, merge, publish, distribute, sublicense and sell the software. It also permits running, deploying and fine-tuning the model, as well as creating derivative works.

Those permissions cover the model weights, parameters, configuration files, inference and training code, and associated documentation. Copies or substantial portions must retain the copyright and permission notice, and use must comply with applicable laws and regulations.

The software, outputs and results are provided without warranty, and the licence includes a limitation-of-liability provision. These are broad permissions, but they are not the unconditional MIT terms attached to GLM-5.2.

The security review has a two-part trigger

The additional security review applies only when both of these conditions are met:

  • The licensee or one of its affiliates operates a Model-as-a-Service business.
  • The licensee and its affiliates exceed US$10 billion in aggregate revenue during any consecutive 12-month period.

When both tests are met, the licence says the licensee must pass Z.ai’s security review before using GLM-5.3 or its derivative works for any commercial purpose.

The licence defines Model as a Service as giving a third party access to model inference or fine-tuning while allowing that party meaningful control over inputs, parameters or training data. It expressly excludes model capabilities embedded only within specific end-user features or harnesses. It also excludes merely relaying requests to models hosted by someone else.

That wording focuses the condition on very large companies that operate model-hosting businesses. The New Stack’s independent review reached the same practical reading, distinguishing model hosts from ordinary embedded products and request routers. Organisations with mixed hosting and product businesses should still assess their exact structure rather than assume an exclusion applies.

The review process remains unspecified

The licence says Z.ai may reasonably determine the scope and method of its security review, but it does not provide a checklist, timetable, evaluation standard, validity period or appeal process. It provides an email address for licence questions.

The New Stack also reported that Z.ai said it held the weights for two additional weeks for safety evaluation and hardening. That timing is a company statement, not an independently audited safety result. The licence itself does not explain which controls the review will test.

For a provider near the revenue threshold, that missing procedure creates a planning question. A team may know that review is required without knowing how long it will take or what evidence Z.ai will request.

What developers and operators should check

The model card lists deployment paths through vLLM, SGLang, Transformers, KTransformers and Unsloth. It also documents reasoning_effort settings of low, high and max. These are implementation options, not evidence that a 753-billion-parameter model will fit a given hardware or cost envelope.

Teams evaluating GLM-5.3 should pin the exact model and licence revision, retain the required notices, test the model within their own infrastructure and confirm whether their business falls within the Model-as-a-Service definition.

Large commercial providers should resolve the licence condition with their legal, security and procurement teams before deployment. The licence may change after publication, so the current text should be checked again before any consequential decision. This article explains the published terms and is not legal advice.