Zhipu AI appears to be testing a faster version of its latest model. A model called GLM-5.3-Highspeed has started showing up in ZCode, the company's coding tool, according to QbitAI, which cited a user who first spotted it in ZCode version 3.15.1.
Access is limited. The model is being rolled out as a gray release to randomly selected subscribers, and it is not available to everyone on a paid plan. A QbitAI reporter on the Lite plan tried to add and switch to the model manually but could not use it, either because that tier is not eligible or because the account had not been included in the test.
The figure drawing attention comes from users rather than Zhipu. A screenshot shared by the user who found the model showed generation at roughly 490 tokens per second. At that pace, producing 10,000 tokens would take about 20 seconds of generation time. Zhipu has not announced the model or published an official throughput figure, so the number should be read as a user measurement, not a guaranteed service level.
For comparison, GLM-5.3-FlashX, which Zhipu released last month, peaked at about 200 tokens per second, while Artificial Analysis measured GLM-5.3-Flash on the Z.ai API at around 60 tokens per second.
There were earlier hints of the model. The model catalog of Pi, an open-source coding agent, already lists GLM-5.3-Highspeed with a context window of 1 million tokens and a maximum output of 131,072 tokens, or about 128k, along with low, high and max reasoning-effort settings.
Users on Reddit have also asked whether the model is connected to Omen Alpha, an anonymous model that appeared on OpenCode in September and has been widely linked to GLM. Guesses about Omen Alpha's identity vary, with some pointing to GLM-5.3-Highspeed and others to an Omni or a 5.4 Flash variant. None of this has been confirmed by Zhipu.
If speeds near 490 tokens per second hold up in wider use, the model would matter most for long-running agent and coding tasks, where waiting on generation adds up. For now it remains a limited test inside ZCode.
SEE ALSO: Zhipu Opens GLM-5.3-FlashX Near 200 Tokens/s on ~100k Domestic Accelerators · GLM-5.3 Arrives Hours After Tang Jie's 'sooooooon'