Z.ai releases GLM-5.3, a 753B open-weights model aimed at coding and cyber tasks

Z.ai released GLM-5.3 on Hugging Face. The card says it uses the same base model as GLM-5.2, so every gain comes from post-training. The Hugging Face record lists about 753 billion parameters in total.

The maker claims it is the most capable open-weights model for coding, with a 50 percent improvement over GLM-5.2 on its in-house Z.ai Code Bench. It also reports gains on Terminal Bench 3.0 and a result it describes as state of the art on CyberGym for vulnerability discovery. These numbers are Z.ai's own, taken from its own table. In that table GLM-5.3 trails Kimi K3 on some rows and leads on others.

Thinking depth is set with reasoning_effort at low, high or max; max is the default. The card lists SGLang, vLLM, Transformers, KTransformers and Unsloth as ways to run it. The licence is listed as glm-5.3, a custom licence rather than a standard one, so read it before building on it.

The card gives no release date; the date used is the Hugging Face repo creation date.

A post-training refresh of a known base. The cyber claims are worth reading twice.

Sources

• Model card on Hugging Face
• Hugging Face repo metadata (creation date, licence, size)

Editor's first pass (AI-assisted)

• Real 4: Ungated weights, with vLLM, SGLang and Transformers listed on the card; the licence is custom.
• Useful 4: Coding and agent benchmarks are the focus; reasoning effort has three levels.
• Notable 4: Over 2,000 likes and about 1.9 million downloads on the API record.
• Novel 2.5: The card says it is the same base model as GLM-5.2; every gain comes from post-training.