Zhipu Tests GLM-5.3-Highspeed in ZCode Gray Release, Users Report ~490 Tokens/s

Zhipu AI is testing GLM-5.3-Highspeed, a faster variant of its latest model, within its coding tool ZCode. The model appeared in ZCode version 3.15.1 as a gray release for randomly selected subscribers, meaning it is not accessible to all paid users. One user measured generation speeds at approximately 490 tokens per second, though Zhipu has not published an official throughput figure.

Performance and Comparison

The reported speed of 490 tokens/s allows for generating 10,000 tokens in about 20 seconds. This represents a significant increase over previous versions:

• GLM-5.3-FlashX: Peaked at about 200 tokens per second.
• GLM-5.3-Flash: Measured at around 60 tokens per second on the Z.ai API by Artificial Analysis.

Technical Specifications and Context

While Zhipu has not officially released the model, the open-source coding agent Pi lists GLM-5.3-Highspeed in its catalog with the following specifications:

• Context window: 1 million tokens
• Maximum output: 131,072 tokens (~128k)
• Reasoning settings: Low, high, and max effort levels

Community speculation links the model to Omen Alpha, an anonymous model that appeared on OpenCode in September, though Zhipu has not confirmed this connection. If the high speeds are maintained, the model would primarily benefit long-running agent and coding tasks where generation latency is a bottleneck.