BeijingZhipu’s fast GLM-5.3 reportedly outputs 490 tokens a second.

Coding agents make dozens of model calls per task, so tokens a second set the wait.

Zhipu’s GLM-5.1 high-speed service hit 400 tokens a second by compiling the whole compute graph ahead of time into one chip kernel.

That is one blogger’s figure from a limited test, against 70.84 for standard GLM-5.3.

Zhipu has not disclosed a price or when the model will serve users at scale.

How each outlet framed it
Zhidx
benchmarks GLM-5.3 HighSpeed at 490 tokens/s against DeepSeek V4.1 Flash at 121 and GLM-5.3 at 71, using AI Ping data

Sources: Zhidx