Zhipu GLM-5.3 Matches Fable 5 On Coding Using Only Post-Training

Zhipu has released GLM-5.3, a large language model that prioritizes software engineering and cybersecurity tasks. This release matters because it demonstrates that significant performance gains in specialized coding and security domains can be achieved through post-training alone. Developers and security researchers can now access a model that matches proprietary competitors without requiring architectural overhauls.

Model identifies legacy vulnerabilities and leads CyberGym benchmarks

The model operates with 743 billion parameters, the same scale as its predecessor GLM-5.2. Zhipu achieved the new performance levels by refining the model through post-training rather than changing the underlying architecture. Z.ai, the entity managing the release, describes the model as built for coding and ready for cyber defense.

GLM-5.3 achieves a score of 66.9 on the DeepSWE benchmark and 28.3 on Terminal Bench 3.0. On the CyberGym benchmark, the model leads with a score of 84.5. These metrics indicate strong agentic capabilities in automated coding and security analysis workflows.

The model identified 2,436 security findings across 269 open-source projects during testing. One of the vulnerabilities discovered dates back to 1981, highlighting the model's ability to detect legacy code issues. Zhipu claims the model matches the coding performance of Fable 5 using only post-training methods.

Zhipu will release the model weights for GLM-5.3 globally in a few days. This open release allows the community to evaluate the model's security and coding capabilities directly. The weights will be available for immediate use by developers interested in agentic AI applications.