Zhipu has commenced grayscale testing for its cutting-edge, high-speed flagship model, GLM-5.3 HighSpeed, within its AI programming tool, ZCode. This model is remarkable for its output speed, which surpasses 490 tokens per second—several times faster than the prevailing mainstream large models in China. It offers a context window of up to one million tokens, with a maximum output capacity of 131,072 tokens, while maintaining the same parameters as the standard GLM-5.3 model. Presently, access to this model is randomly granted to ZCode subscribers, with no explicit internal testing application portal available and pricing details still pending announcement. Previously, Zhipu introduced the GLM-5.1 High-Speed Version, which reached an output speed of 400 tokens per second. The exceptional speed of GLM-5.3 HighSpeed can be attributed to systematic optimizations implemented across three key levels: the inference engine, scheduling system, and underlying infrastructure. These enhancements were achieved through collaborative efforts between the Zhipu team and the TileRT team. Currently, accelerating the inference speed of flagship models has emerged as a new competitive frontier among large model providers.