The release keeps the GLM-5.2 base model and attributes its gains to scaled post-training. Z.ai expanded executable environments, synthesized verifiers, used asynchronous reinforcement learning, and retained SAO compaction for long tasks.
GLM-5.3 is available through Z.ai products and coding plans, with weights announced for a later public release after safety hardening. It supports enabled thinking with low, high, and max reasoning effort, making it useful for coding agents that need adjustable depth and persistent execution.
