From the archive
Z.ai releases GLM-4.5 and GLM-4.5-Air
Two hybrid reasoning models offer hosted access and downloadable weights, with thinking and quick-response modes.
Z.ai introduced GLM-4.5 and GLM-4.5-Air, model families combining reasoning, coding, and tool use. Both arrived through the Z.ai interface and application programming interface (API), with weights available on Hugging Face and ModelScope.
Choose the model and its response mode
GLM-4.5 has 355 billion total parameters, with 32 billion active during an individual computation. The smaller Air variant has 106 billion total parameters and 12 billion active. Both use a mixture-of-experts design.
The release provides a thinking mode for more involved reasoning and tool work, alongside a non-thinking mode for quicker responses. Applications can choose the mode around the kind of task being performed.
Hosted or self-managed access
The downloadable route gives teams an option to run the models themselves, while hosted requests avoid managing that inference environment. The Z.ai profile covers its model endpoints, coding plans, and usage-based access.