Zhipu quietly begins testing GLM-5.3 HighSpeed in ZCode

On October 12, 2026, Chinese AI company Zhipu began quietly rolling out a grayscale test of its high-speed flagship model, GLM-5.3 HighSpeed, within its AI coding tool ZCode. The test is being offered randomly to ZCode subscribers, with no formal application channel.

AI programming blogger Leo reported on X that he has been granted access to the model and that it can generate more than 490 tokens per second. By comparison, AI Ping data shows DeepSeek V4.1 Flash's official API outputs at 121.34 tokens per second, while Zhipu's GLM-5.3 API runs at 70.84 tokens per second. The new model features a 1 million token context window and can output up to 131,072 tokens in a single response, matching GLM-5.3's specifications. Pricing for GLM-5.3 HighSpeed has not been disclosed.

Zhipu has previously released high-speed variants of its flagship models. In May 2026, it launched GLM-5.1-highspeed, an API designed for select enterprise users with an output speed of 400 tokens per second. The company highlighted that high-speed models can meaningfully reduce the time spent on coding agent tasks that require dozens of model calls. Earlier reports indicated that the GLM-5.1-highspeed API was built with the TileRT team through system-level optimization of the inference engine, scheduling system and infrastructure.

GLM-5.3 HighSpeed was also listed on the website of open-source coding agent Pi harness about a month before this grayscale test, in September 2026. Improving model inference speed has become a growing focus among AI vendors, with several moving beyond dedicated fast models to push the speed of their flagship models to the next level.