GLM 4.5 Air

GLM 4.5 Air is the smaller sibling in Z.ai's GLM-4.5 family of agent-oriented foundation models, sharing the same Mixture-of-Experts lineage as the flagship. It adopts a streamlined design with 106 billion total parameters and 12 billion active per forward pass, balancing capacity against per-query compute. Z.ai positions the variant, alongside the larger GLM-4.5, as foundational for applications that lean on tool invocation, web browsing, software engineering, and front-end development, with documented integration paths into code-centric agents such as Claude Code and Roo Code. Its intended use is therefore less about open-ended chat and more about driving multi-step, tool-using workflows where structured reasoning and reliable API calls matter.

The model is trained on the same pipeline as the rest of the GLM-4.5 series: an initial pretraining phase on roughly 15 trillion tokens of general-domain data, followed by targeted fine-tuning on code, reasoning, and agent-specific datasets, and a reinforcement learning stage focused on reasoning, coding, and agent performance. Z.ai evaluates the family across 12 representative benchmarks in reasoning, coding, and agentic ability, framing the GLM-4.5 release as a state-of-the-art open-source MoE model. A hybrid reasoning setup lets callers toggle between a Thinking mode for complex tool-using tasks and a Non-Thinking mode for quicker responses, with dynamic thinking enabled by default. For practitioners, this makes GLM 4.5 Air a practical fit when an agent harness needs compact reasoning, code generation, and tool-calling behavior in one model rather than a separate planner and executor.