# GLM-5.3 / GLM-5.2 / GLM-5.1 / GLM-5

A family of open-weight language models from Z.ai focused on coding and long-horizon agentic tasks, with GLM-5.3 adding stronger coding and emergent cyber capability.

GLM-5.3 uses the same base model as its predecessor GLM-5.2, with every capability gain coming from post-training; compared with GLM-5.2 it is described as much stronger at complex coding and long-horizon tasks, achieving open-source state-of-the-art results on public benchmarks including Terminal Bench 3.0 and Agents' Last Exam, and showing emergent cyber capability that the developers say grew faster than expected during scaled post-training, reaching state-of-the-art results on the CyberGym vulnerability-discovery benchmark.

A separate variant, GLM-5.3-Flash, starts from a newly trained base model with its architecture and training recipe redesigned around efficiency, introducing a hybrid sparse-and-linear attention architecture that reduces long-context serving cost while preserving long-context accuracy, plus a technique called Manifold-Constrained Hyper-Connections and a 30-trillion-token multimodal pre-training corpus.

The earlier GLM-5.2 release introduced a solid 1M-token context window that stays stable for long-horizon work, flexible coding effort levels to trade off performance and latency, and an architecture change called IndexShare that lowers per-token compute at long context lengths.

Website: https://z.ai/blog/glm-5.2