Z.ai Releases GLM-5.3 Under Conditional License
Z.ai put the GLM-5.3 model weights on Hugging Face under licensing terms that require providers earning more than $10 billion in revenue to undergo a security review. Z.ai's GitHub documentation describes GLM-5.3 as sharing GLM-5.2's base model, with reported gains attributed to post-training, while the related GLM-5.3-Flash release is listed under the MIT license.
Z.ai released GLM-5.3 model weights on Hugging Face with licensing terms that differ from the MIT license listed for its related GLM-5.3-Flash release. Providers with more than $10 billion in revenue must complete a security review under the new GLM-5.3 license.
The license condition is consequential because it distinguishes open-weight availability from unrestricted commercial reuse. The original report describes the requirement as aimed at hyperscale providers, though the available source material does not provide the full license text or details of the review process.
Model lineage and claimed capabilities
Z.ai's GitHub repository states that GLM-5.3 uses the same base model as GLM-5.2 and attributes its improvements to post-training. The repository reports stronger results on complex coding and long-horizon tasks, including a 50% improvement on Z.ai's in-house Code Bench, as well as claimed state-of-the-art results among open-source models on Terminal Bench 3.0 and Agents' Last Exam.
The repository also reports that cyber capability improved during post-training, with GLM-5.3 reaching state-of-the-art performance on CyberGym for vulnerability discovery. Those are vendor-reported benchmark claims, and practitioners should evaluate them against independent testing and their own task distributions.
A separate MIT-licensed Flash release
The Hugging Face page for GLM-5.3-Flash currently lists an MIT license. That model is described by Z.ai as a natively multimodal mixture-of-experts system with 320 billion total parameters and 18 billion active parameters. Its architecture combines sparse and linear attention, according to the model card, and the repository lists support for local deployment through several serving frameworks.
The differing terms matter operationally. Organizations assessing open-weight models generally need to separate checkpoint access, serving requirements, redistribution rights, and provider-specific contractual conditions. In this case, teams evaluating GLM-5.3 should review the model-specific license rather than infer its permissions from the MIT designation shown for GLM-5.3-Flash.
The available material does not specify which security-review criteria apply, how approval is administered, or whether the requirement affects downstream hosting and fine-tuning. Those details are likely to determine the practical reach of the restriction for cloud and model-serving providers.
Key Points
- 1GLM-5.3 weights are publicly available, but providers above $10 billion in revenue face a security-review condition.
- 2Z.ai's repository attributes GLM-5.3 improvements to post-training rather than a newly trained base model, focusing evaluation attention on alignment and agentic capabilities.
- 3Separate GLM-5.3 and GLM-5.3-Flash licensing illustrates why open-weight adopters should verify terms per checkpoint, not by model-family branding.
Scoring Rationale
The release makes a coding and agentic model available as weights, which is directly relevant to model evaluation and self-hosting decisions. The reported revenue-based security-review condition also introduces a notable licensing constraint for large AI infrastructure providers.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
