GLM-4.5-Air is Z.AI's compact counterpart in the GLM-4.5 family, built from the ground up as a foundational model for agent-oriented applications. It shares its training lineage with the larger flagship: a broad pre-training phase followed by targeted fine-tuning on code, reasoning, and agent-specific datasets, with reinforcement learning layered on to sharpen those capabilities. The result is a model aimed squarely at workflows where the system has to plan, call tools, browse, write code, and chain together multi-step tasks rather than just answer isolated questions.

Architecturally, GLM-4.5-Air adopts a Mixture-of-Experts design with 106B total parameters and 12B active parameters per forward pass, trading some raw capacity for noticeably leaner inference. It runs in two modes, a thinking mode for complex reasoning and tool use, and a non-thinking mode for faster, more conversational responses, so developers can dial effort up or down depending on the task. Optimizations target tool invocation, web browsing, software engineering, and front-end development, making it a natural fit for code-centric agents like Claude Code and Roo Code as well as custom agent pipelines built through tool-calling APIs. Weights are openly published on Hugging Face under the zai-org namespace, which simplifies self-hosting, fine-tuning, and integration into private stacks.