GLM-5.3 is an open-weights release from the GLM family that keeps the same base model as GLM-5.2 and channels all of its improvements into scaled post-training. The stack that produced those gains includes the IndexShare approach for efficient long-context processing, the SAO method for reinforcement learning on long-horizon tasks, and the slime framework for large-scale asynchronous training, all run on a growing library of long-horizon task environments. By leaning on more environments, more diverse tasks, and more compute rather than a new foundation, the release is positioned as a deliberate post-training push aimed at complex software work and long-running agentic workflows.

The post-training focus shows up clearly in the model’s benchmark profile, where it reaches open-source state-of-the-art on Terminal Bench 3.0 and Agents’ Last Exam, posts a roughly 50 percent gain over its predecessor on the in-house Z.ai Code Bench, and leads the field on CyberGym for vulnerability discovery. The exploitation results are especially striking, more than doubling GLM-5.2 on ExploitGym and posting strong scores on ExploitBench, with the largest jumps appearing further up the exploitation chain. Strong showings on DeepSWE v1.1, NL2Repo, FrontierSWE, SWE-Marathon, and ProgramBench reinforce its fit for coding assistants, repository-level engineering tasks, and agentic pipelines that need to plan, code, and reason across long contexts.