Chinese AI company Zhipu has officially open-sourced GLM-5.3, its 744B-parameter flagship model, with FP8 and BF16 weights now live on Hugging Face. Unlike last generation’s GLM-5.2, this release doesn’t use MIT — it ships under a new “GLM-5.3 License.”
The new license stays fairly permissive for everyday developers: you can use, modify, distribute, deploy, fine-tune, and even sell the model and its derivatives. But Zhipu added a restriction aimed squarely at big model-as-a-service companies. If your company and its affiliates pull in more than $10 billion in total revenue over any consecutive 12-month period, you’ll need to pass Zhipu’s security review before using GLM-5.3 commercially. Ordinary products that embed the model into specific features are exempt.
GLM-5.3 was actually released two weeks ago. Zhipu held back the weights because the model’s cybersecurity capabilities grew faster than expected, so it wanted to run extra safety evaluations and hardening first. The model now supports self-hosting via SGLang, vLLM, Transformers, KTransformers, and other frameworks.
