GLM-5
Zhipu's frontier open-weight model, trained entirely on domestic Huawei hardware
GLM-5 is Zhipu AI’s frontier open-weight model, released February 11, 2026 as roughly a two-times scale-up from GLM-4.5, growing from 355 billion to about 744 billion total parameters while keeping a similar 40 billion active per token in its mixture-of-experts design. It ships with a 203,000 token context window, well short of the 1 million tokens its GLM-5.2 successor would add four months later, and Zhipu trained it entirely on Huawei Ascend chips using the MindSpore framework rather than Nvidia hardware. API pricing starts at $0.57 per million input tokens and $1.92 per million output tokens through Zhipu’s own platform and the WaveSpeed API. It shipped under the MIT license, and coverage at the time framed it as the first frontier-class model from a Chinese lab built on fully domestic hardware.