GLM-5.3 went open weight under a custom licence in place of MIT

Z.ai released the weights for its flagship GLM-5.3 model on Hugging Face on Friday 28 August. Frederic Lardinois reported for The New Stack that the licence changed at the same time.

GLM-5.2 was released under MIT. GLM-5.3 is released under something Z.ai calls the GLM-5.3 licence.

What the licence says

For an individual, the practical position is unchanged. The licence grants the right to run, deploy and fine-tune.

The new clause targets companies that host the model, as opposed to routing it the way OpenRouter does or embedding it in a product. Any such company with aggregate revenue above $10 billion over any 12 consecutive months "must pass Z.AI's security review before using the Software or its derivative works for any commercial purpose."

The New Stack notes that the licence contains no acceptable-use section and says nothing about cyber or offensive security, despite the security framing around the release.

That framing exists because Z.ai says it held the weights back for two weeks of safety evaluation and hardening, which is new: GLM-5.2's weights were available on launch day. The company reports 84.5 percent on CyberGym, a vulnerability discovery benchmark, and says it is the best published result. It also says it used the model to find 2,436 vulnerabilities across 269 open-source projects including the Linux kernel.

Both of those are Z.ai's own figures. The New Stack states plainly that nobody outside the company has reproduced the benchmark, and that only a few dozen of the vulnerability findings are publicly inspectable so far. Treat them as claims until that changes.

How it compares with other labs

Moonshot requires a model-as-a-service provider with more than 100 million active users or $20 million in monthly revenue to display "Kimi K3" prominently in the interface. DeepSeek still uses plain MIT for its flagship models.

Z.ai kept MIT for GLM-5.3-Flash, so it has not abandoned permissive licensing. It has stopped applying it to its best model.

Running it

Same architecture as GLM-5.2: 753 billion parameters, mixture of experts, a 1 million token context window and a maximum output of 128,000 tokens. Weights are released in BF16 and FP8, and run on vLLM, SGLang, KTransformers and Transformers.

Local deployment is not practical for most people. Unsloth's 2-bit quantisation, which it says keeps about 86 percent top-1 accuracy, needs 245GB of memory. That fits a Mac with 256GB of unified memory and nothing smaller. The 8-bit version needs 810GB.

Hosted, it is $1.40 per million input tokens and $4.40 per million output.

What to actually do

If your company is far below $10 billion in revenue, this changes nothing for you today, and the model is cheap and capable.

If you are building a product on top of it, read the licence yourself rather than a summary of it, including this one. Bespoke model licences are new enough that the phrase "open weight" now covers several genuinely different legal positions, and the difference only becomes visible when someone asks who is hosting the inference.

Source

• Z.ai's GLM-5.3 goes open weight, but its new license aims at hyperscalers, The New Stack, 28 August 2026