Zhipu appears to be testing a faster version of its latest model. The model, named GLM-5.3-Highspeed, has started appearing in ZCode, the company's coding tool, according to a report by QbitAI. The source of this information is a user who first discovered the model in version ZCode 3.15.1.
Access to this model is restricted. It is being rolled out as a beta release for randomly selected subscribers and is unavailable to all paid plan users. One QbitAI reporter using the Lite plan attempted to manually add and switch to this model but was unsuccessful, possibly because their access level is unsuitable or their account was not included in the test.
The figures that attracted attention come from users, not from Zhipu itself. The user who discovered the model shared a screenshot showing that generation occurs at a speed of approximately 490 tokens per second. At this speed, generating 10,000 tokens would take about 20 seconds. Since Zhipu has not made an official announcement about this model nor published official throughput metrics, this number should be considered a user-generated measurement rather than a guaranteed service level.
For comparison, GLM-5.3-FlashX, released by Zhipu last month, achieved a peak speed of around 200 tokens per second, while Artificial Analysis measured GLM-5.3-Flash via the Z.ai API at a speed of approximately 60 tokens per second.
There were previous hints about this model. The Pi models catalog, an open coding agent, already includes GLM-5.3-Highspeed with a context window of 1 million tokens and a maximum output of 131,072 tokens, or about 128 thousand, as well as low, high, and maximum reasoning effort settings.
Users on Reddit have also questioned the connection between this model and Omen Alpha, an anonymous model that appeared in OpenCode in September and was often linked to GLM. Speculations about Omen Alpha's identity vary: some point to GLM-5.3-Highspeed, others to Omni or a 5.4 Flash variant. None of this has been confirmed by Zhipu.
If the speed near 490 tokens per second is maintained with wider adoption, this model will be most significant for long-running agent and coding tasks where generation waiting time accumulates. For now, it remains a limited test within ZCode.