Zhipu quietly tests GLM-5.3-Highspeed at 490 tokens per second
Zhipu is reportedly gray-testing GLM-5.3-Highspeed inside Zcode, promising unusually fast generation but offering no public model release yet.
Zhipu appears to be quietly testing a new GLM-5.3-Highspeed model in its Zcode coding product. QbitAI reported that selected subscribers could access the model through a gray rollout, while one user measurement showed generation approaching 490 tokens per second.
A speed claim, not a launch
The reported throughput would be a major jump over Zhipu’s recently introduced GLM-5.3-FlashX, which was advertised at up to 200 tokens per second. At 490 tokens per second, generating 10,000 tokens would take roughly 20 seconds once the response stream begins, potentially changing the feel of long coding and agentic tasks.
The evidence remains limited. The figure comes from a user screenshot rather than a standardized public benchmark, and QbitAI was unable to obtain access through its own account. The model is reportedly available only to randomly selected Zcode users, with no public weights, API documentation, pricing, or reproducible evaluation disclosed.
Directory listings suggest a one-million-token context window, a maximum output of 131,072 tokens, and selectable low, high, and maximum reasoning modes. Those specifications appear similar to the regular GLM-5.3 listing, so they should be treated as provisional rather than confirmed product documentation.
Why the rollout matters
If sustained outside a narrow demo path, high generation speed could lower latency for coding agents that repeatedly call a model, especially when responses are long but reasoning quality is already adequate. It could also pressure rival providers to compete on time-to-useful-output rather than only benchmark scores.
For now, the important fact is the uncertainty: this is a gray-tested service configuration, not a verified model release. Until Zhipu publishes official access terms and independent measurements, the 490-token figure is a signal of engineering ambition, not yet a dependable market capability.