Zhipu AI Tests GLM-5.3-Highspeed Model With Reported 490 TPS in Limited Grayscale Rollout

THE ESSENTIALS

GLM-5.3-Highspeed is in limited grayscale testing within Zcode 3.15.1, randomly assigned to select subscribers.
A user reported output speeds of 490 tokens per second, compared to 200 TPS for GLM-5.3-FlashX and roughly 60 TPS for GLM-5.3-Flash.
The model supports a 1-million-token context window and up to 128K output tokens with three thinking-intensity tiers.
Zhipu previously launched GLM-5.1-highspeed in May at 400 TPS; the new variant extends that speed advantage further.

Grayscale Testing Underway

Zhipu AI has started limited grayscale testing of a new model variant called GLM-5.3-Highspeed within its Zcode coding assistant platform. According to a report by QbitAI published on October 12, the model is currently available only to randomly selected subscribers using Zcode version 3.15.1 and is not yet generally accessible.

The rollout appears restricted, with even some paying users on Lite-tier plans unable to access the model. Zhipu has not issued an official announcement regarding availability timelines or pricing for the new variant.

Reported Speed Gains

A user identified as @leo114119 shared test results showing GLM-5.3-Highspeed reaching output speeds of 490 tokens per second. If sustained, that pace would generate 10,000 tokens in roughly 20 seconds during the generation phase alone.

The reported figure represents a significant jump over existing Zhipu models. GLM-5.3-FlashX, released last month, tops out at approximately 200 tokens per second, while GLM-5.3-Flash on the Z.ai API delivers around 60 tokens per second according to Artificial Analysis benchmarks. Zhipu's previous high-speed model, GLM-5.1-highspeed, launched in May at 400 tokens per second.

Model Specifications

According to model directory listings, GLM-5.3-Highspeed supports a context window of 1 million tokens and a maximum output of 131,072 tokens (approximately 128K). It offers three configurable thinking-intensity levels — low, high, and max — matching the configuration options available on the standard GLM-5.3 model.

Community discussion has also linked the new variant to an anonymous model called Omen Alpha that appeared on OpenCode in September, though Zhipu has not confirmed any connection. Users who have tested the model describe it as fast, capable of extended reasoning, and competitive in output quality.