Zhipu AI releases GLM-5.3 under an open-weight license

Zhipu AI just dropped GLM-5.3 under an open-weight license, giving developers a top-tier coding and cyber-security model to run locally.

Zhipu AI released GLM-5.3, built on the exact same base model as GLM-5.2. According to Zhipu AI, every gain comes entirely from post-training. The new version scores a 50% boost over GLM-5.2 on their internal Z.ai Code Bench, hits 28.3 on Terminal Bench 3.0, and more than doubles its predecessor's performance on security benchmarks like ExploitBench.

Why it matters: It sets a high mark for open-weights models in complex coding and long-horizon tasks. Developers get local access to an agent capable of advanced vulnerability discovery and software development without renting proprietary APIs.

Here's the gist on running it: you can deploy GLM-5.3 using frameworks like SGLang, vLLM, KTransformers, and Unsloth, or on Ascend NPUs with vLLM-Ascend. You can adjust its reasoning budget with the reasoning_effort parameter (low, high, or max). For chat setups, make sure to pass clear_thinking=true in the chat template.

Local post-training gains strike again—now your terminal can trace exploits while you grab coffee.