China’s Z.ai announced the weights of the GLM-5.3 artificial intelligence model on the Hugging Face platform, but refused MIT’s democratic permission. Service providers with more than $10 billion in revenue are now subject to developer security reviews.
Photo credit: Milad Fakurian / unsplash.com
A notable feature of the Z.ai GLM-5.3 model is its post-training mechanism, which allows it to significantly improve the quality of work over GLM-5.2 and be on par with leading US models. At the same time, the developers changed the license: GLM-5.2 belongs to the standard MIT, while GLM-5.3 has a special license. Companies that want to launch the product and have revenue exceeding $10 billion for 12 consecutive months “Before using the software or its derivatives for any commercial purpose, it must pass Z.ai’s security review”.
The tightening of the GLM-5.3 license looks like a clear departure from China’s current practice of distributing open artificial intelligence models as freely as possible, but that doesn’t mean Z.ai is giving up on it. For most developers, the conditions remain almost the same: models are still allowed to be freely run, modified, further trained and used in commercial products. The restriction is primarily aimed at the largest cloud and artificial intelligence vendors who hope to make money by offering access to GLM-5.3 as a standalone service. In this way, Z.ai seeks to combine widespread adoption of its technology with control by major players over its commercial use and associated security risks.
Although the GLM-5.2 weights were released on the day of the announcement, the developers needed two weeks between the announcement and the release of the open weights to evaluate safety; the developers highlighted that GLM-5.3 scored 84.5% in the CyberGym test, the best result of all the tests. Z.ai used it to analyze 269 open source software projects, including the Linux core, and found 2,436 vulnerabilities; however, not all results are available to the public. GLM-5.3 features a hybrid expert architecture, 753 billion parameters, a context window of 1 million tokens, and a maximum output of 128,000 tokens. Scales are available in BF16 and FP8 formats; the model is run using vLLM, SGLang, Ktransformers and Hugging Face Transformers.
Running Z.ai GLM-5.3 on consumer hardware will be difficult: even with 2-bit quantization with 86% accuracy, it requires 245 GB of memory, while 8-bit requires 810 GB. When accessing the model via the API, it costs $1.40 (1 million input tokens) and $4.40 (1 million output tokens) to use the model. Z.ai also revealed the weight of the GLM-5.3-Flash, which was first unveiled as the Anonymous Ox Alpha. It sells for $0.15/$0.47 respectively, runs on the China Artificial Intelligence Accelerator, and is available under the traditional MIT license.
So far, openness has been one of the main competitive advantages of Chinese AI developers. DeepSeek, Alibaba, Moonshot AI and Z.ai itself have been allocating weight to cutting-edge models, allowing third-party companies to build their own products on top of them for free, rapidly expanding China’s AI ecosystem globally. The new GLM-5.3 license demonstrates that as such models move ever closer to the cutting edge, developers are beginning to seek trade-offs between openness, security and the ability to benefit from the technology they create.
If you find an error, select it with your mouse and press CTRL+ENTER.
