The GLM-5.3 License is the bespoke licence under which Z.ai released the weights of GLM-5.3 in August 2026. It is published as the LICENSE file at the root of the zai-org/GLM-5.3 Hugging Face repository, in parallel English and Chinese texts whose operative terms match, and carries a 2026 Z.AI copyright notice (Source: huggingface.co). It replaces the unrestricted MIT licence under which GLM-5.2 and the GLM-5 and GLM-4.5 generations were released.
Structurally the licence is the MIT licence with one clause added. Clauses 1 and 3 reproduce MIT's attribution requirement and its warranty and liability disclaimer nearly verbatim, with a sentence added to clause 1 requiring that the licensee's use comply with applicable laws and regulations. The departure from MIT is clause 2 alone.
Grant
The grant is broad and is stated before the conditions. It runs to "any person or entity" obtaining a copy of the Software, which the licence defines as "the model weights, parameters, configuration files, inference and training code, and associated documentation." The permitted dealings are unrestricted and enumerated: use, copy, modify, merge, publish, distribute, sublicense and sell copies; run, deploy, fine-tune or otherwise modify the Software and create derivative works from it; and permit others to do the same.
The security-review condition
Clause 2 defines "Model as a Service" (MaaS) functionally rather than by product category: giving a third party access to language model inference or fine-tuning, for example via API, "in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data." Two exclusions are stated — end-user products where model capabilities are "solely embedded within specific features or harnesses," and "mere relaying of requests to models hosted by others."
The condition itself has two limbs, both reaching corporate groups rather than single entities:
- The trigger applies where "the Licensee or any of its affiliates operates a Model as a Service business"; and
- the revenue test runs on "the aggregate revenue of the Licensee and its affiliates," exceeding US$10 billion or its currency equivalent "in total over any consecutive 12 months."
Where both are met, the licensee "must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose." The obligation is therefore broader than the trigger: the MaaS test determines who is caught, but once caught the bar covers all commercial use, not only hosting the model as a service.
The licence does not define the review. It states only that "the scope and method of the security review shall be reasonably determined by Z.AI," leaving the gate unilateral and its content undisclosed, with "reasonably" the sole constraint. No timetable, appeal, standard of approval, or fee is specified. Licence questions are directed to glmlicense@z.ai.
Reception
Frederic Lardinois, writing in The New Stack on 28 August 2026, characterised the terms as more restrictive than those used by other Chinese laboratories, comparing them with the licence for Moonshot's Kimi K3, which requires only that a model-as-a-service provider above 100 million active users or $20 million in monthly revenue display the model name prominently in the product interface (Source: thenewstack.io).
Because the weights are public, the condition is not technically enforceable against a party willing to breach it; it operates as a contractual term binding the largest hosts rather than as a distribution control. The $10 billion group-revenue threshold reaches hyperscalers and large cloud providers while leaving smaller deployers under the MIT-equivalent terms.
Relationships
- supersedes: the MIT licence used for GLM-5.2 and earlier GLM releases.
- related: GLM-5.3, Zhipu AI, Open-Weight Frontier Models, Open-Source AI / Open-Weight Models, Kimi K3