GLM-5.3's Open Licence Adds a Security Review for $10bn Rivals

Z.ai published the weights of the full GLM-5.3 on 28 August, following the Flash variant released a day earlier. The release is being described as open source. The licence file in the repository — three thousand characters, readable in a minute — contains a clause that most open-weight licences do not.

The clause

The licence defines "Model as a Service" as giving a third party access to inference or fine-tuning in a way that lets them control inputs, parameters or training data. It then states that if a licensee and its affiliates run such a business and their aggregate revenue exceeds $10 billion over any consecutive 12 months, the licensee "must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose." The next sentence is the operative one: "The scope and method of the security review shall be reasonably determined by Z.AI."

Who that describes

The threshold is drawn at a level that excludes essentially every startup and research group and captures the large cloud and model providers. Two carve-outs narrow it further: end-user products that merely embed model capabilities are excluded, as is the mere relaying of requests to models hosted by others. The clause is aimed squarely at rivals who would serve GLM-5.3 as a paid API.

What the common framing gets wrong

"Open source" implies use without permission. Here the largest class of commercial users needs an approval whose criteria are written by the licensor after the fact, and the licensor is a direct competitor in the same market. That is closer to a field-of-use restriction with a discretionary gate than to an open licence. The weights are genuinely downloadable and the model genuinely runs — the restriction is on who may resell it, and it is not a formality.

The benchmark to discount

The coding improvement leading the coverage is measured on Z.ai's own private code benchmark, not a public suite. Its strongest published security result is a single row on one evaluation; on the exploitation task in the same table, Z.ai's own figures put a competing frontier model well ahead. Neither point is hidden — both are on the model card, under the headline.