According to CoinMeta, Zhipu has officially open-sourced its flagship models with 744b parameters, GLM-5.3, fp8, and bf16. The weights for these models were released alongside Hugging Face. Unlike the previous generation GLM-5.2, this time MIT is no longer used; instead, a new combination of „GLM-5.3 and license“ has been adopted. The new license terms are still relatively lenient towards ordinary developers, allowing them to use, modify, distribute, deploy, fine-tune, and even sell the models and their derivative versions. However, Zhipu has added a restriction specifically for large „model-as-a-service“ companies: if such companies and their affiliates have total annual revenue exceeding $10 billion for any consecutive 12 months, they must first pass Zhipu’s security review before using GLM-5.3 for commercial purposes. Ordinary products that embed the models into specific functions are not subject to this restriction. GLM-5.3 was released two weeks ago, but due to the model’s security capabilities exceeding expectations, Zhipu decided to conduct additional security assessments and enhancements before making the weights available. Now, the model also supports deployment in frameworks such as sglang, vllm, transformers, and ktransformers.