The Chinese laboratory publishes an update obtained without retraining the basic model, with 743 billion parameters and a context of one million tokens.
Z.ai posted GLM-5.3 on August 14, 2026, presented as its most effective model for code and long-term agent tasks. Particularity of this version: the basic model remains that of GLM-5.2, all the gains from a considerably extended post-training.
On paper, the model has 743 billion parameters, accepts a context window of one million chips and produces up to 128,000 chips in output. Z.ai insists on a point often overlooked in comparisons of reasoning models: GLM-5.3 would achieve better results than its predecessor while consuming significantly fewer chips per task, which changes the economic equation of agents that rotate several hours.
The model is accessible via the Z.ai API and the GLM Coding Plan subscription, as well as in the ZCode environment; The publication of weights is announced after a safety review. The company also states that the cybersecurity model's capabilities have progressed faster than it anticipated as training progressed on a scale—a formulation that, coming from a laboratory, deserves to be followed.
Read the full article on the original source
Go to the source →