What changed when Z.ai released the GLM-5.3 weights, and why is the licence no longer MIT?

Yesterday, Z.ai put the weights of GLM-5.3 on Hugging Face. Weights are the learned numbers of a model, and publishing them lets anyone download and run it. The model itself launched as an API earlier in August, so this is the second half of a release that Z.ai split in two.

Two things are new for this lab. It held back the weights for safety testing, and it dropped the MIT licence it had used for its recent flagships. GLM-5.2, released in June, was MIT and open on launch day.

What is in the release

Underneath, this is the base model of GLM-5.2. Z.ai says every gain comes from post-training, the stage after basic pre-training in which a model learns by supervised examples and reinforcement learning. That is unusual, since it suggests a lab can move a model forward a long way without retraining it from nothing.

Both are mixture-of-experts models, in which only some parameters work on each token (a small piece of text). The active count drives compute per token, while the total drives the memory you need.

How it was held back, and why

Z.ai describes the model's cyber capability as having grown faster than it expected as post-training scaled. On its own table, GLM-5.3 scores 84.5 on CyberGym, a vulnerability-discovery test, against 77.2 for GLM-5.2. On ExploitBench, it scores 54.4 against 24.4. Closed models lead that second test: 78.0 for Claude Fable 5 and 76.5 for GPT-5.6 Sol.

The company says it spent two additional weeks on safety evaluation before publishing. The New Stack points out that the numbers are self-reported and that nobody outside the company has reproduced the CyberGym score so far. Z.ai also claims 2,436 vulnerabilities found across 269 open-source projects, with only a few dozen publicly inspectable.

There is also a tension in the licence. The New Stack notes that it has no acceptable-use section and does not mention cyber or offensive security at all. So the safety hold happened before release, and the licence does not then limit misuse. Once weights are public, anyone can fine-tune them, so training-time safeguards can be weakened; that is a general limit of open weights.

Coding results are close to the leaders on Z.ai's table:

What it takes to run

Z.ai lists SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth, along with routes for Huawei's Ascend chips. Let's do the sums for a worked example. The 8-bit files are about 756 GB. On 80 GB graphics cards, that is about ten cards for the weights alone, and working memory comes on top.

The New Stack cites Unsloth's reduced versions: about 245 GB for a 2-bit build, which only just fits a Mac with 256 GB of unified memory, and 810 GB for 8-bit. Quantization, storing each number in fewer bits, saves memory but can cost quality, and a 2-bit build is a heavy cut. Unsloth says the 2-bit build keeps about 86 percent top-1 accuracy, a claim to test on your own tasks.

For most readers, hosted access is the practical route. The New Stack lists Z.ai's own prices at $1.40 US per million input tokens and $4.40 per million output tokens, and $0.15 and $0.47 for Flash. Several other providers already host it, under their own data terms.

Does the GLM-5.3 licence affect you?

Almost certainly not, unless you are a very large host. The licence opens with a broad grant covering use, modification, fine-tuning, deployment and sale. Only one extra condition stands out.

The review's scope and method are for Z.ai to determine "reasonably", in the licence's word, and nothing says what it includes. The threshold counts the whole group, so a cloud giant is covered even if its GLM business is small. If you want no such condition at all, GLM-5.3-Flash remains MIT.

The licence is not the only legal question. Z.ai has been on the US Commerce Department's Entity List, a trade blacklist, since January 2025. What that means for downloading a model is a question for your own lawyer, especially if your organization works under US trade rules. The Canadian position may differ, so get advice before building something serious on it.