Zhipu has launched GLM-5.3-Highspeed, its high-speed large model, on the Zcode platform, with the model entering grayscale testing in Zcode version 3.15.1 and subscription users randomly granted access, iiMedia reported on October 12.
In third-party testing, the model's maximum generation speed reached 490 tokens per second. It offers a 1-million-token context window and a maximum generation cap of approximately 128,000 tokens, with three configurable thinking-intensity tiers: low, high, and max. Compared with the previous generation GLM-5.3-FlashX, whose top speed was around 200 tokens per second, the new model delivers a substantial improvement in generation performance and significantly enhanced long-text output efficiency.
The launch is part of a broader wave of new AI products announced on October 12. Alibaba's Tongyi released the high-speed image model Qwen-Image-2.1-Turbo, supporting 8-step denoising for image generation and editing, with Pro and Turbo APIs going live on the Qwen AI platform on October 11 for developers. Alibaba also released QwenGyre, an online reinforcement learning framework for ultra-long-horizon agent tasks, on the same day.
Industry watchers say 2026 has been defined by the industry as the first year of AI agents, with model speed, context length, and agentic execution capability emerging as the new competitive battleground for Chinese foundation-model developers.
