Z.ai released the weights for GLM-5.3 on Friday. The release package includes 756 GB of model files across 141 Safetensors shards, support for vLLM and SGLang, and an architectural configuration with 256 routed experts where 8 activate per token across a one-million-token context window.
Two technical details in the release notes tell a practical story about where frontier model development is heading.
The first detail is that Z.ai kept the exact base foundation from GLM-5.2. The company did not run a larger pretraining cluster or change the underlying parameter structure. The performance jump came from post-training: reinforcement learning environments, task verification harnesses, and domain data for software engineering and vulnerability hunting. Z.ai reports that Terminal-Bench 3.0 moved from 4.6 to 28.3, DeepSWE went from 46.2 to 66.9, and ExploitBench jumped from 24.4 to 54.4.
Those numbers come from vendor-run evaluations under specific harness configurations, including 400,000 tokens of context and 10-hour timeouts per rollout. Independent reproduction across different serving stacks will show the real operational variance. Still, the architectural choice is clear: pretraining gave them a stable base, and the agentic capabilities were built almost entirely during post-training rollouts.
