GLM 4.5 Air

GLM-4.5-Air is a compact yet powerful hybrid reasoning model purpose-built for intelligent agent applications. It leverages a Mixture-of-Experts architecture that keeps the model efficient by activating only a fraction of its total parameters during inference. The defining feature is its dual-mode operation: thinking mode handles complex reasoning chains and tool usage, while non-thinking mode delivers immediate responses for straightforward tasks. This flexibility makes it well-suited for workflows that mix routine interactions with demanding problem-solving, and it integrates naturally into code-centric agent frameworks that need to plan, browse, and execute tasks.

The training pipeline reflects a deliberate progression from broad knowledge to specialized capability. Starting with pretraining on 15 trillion tokens of general-domain data, the model then underwent targeted fine-tuning on code, reasoning, and agent-specific tasks. Reinforcement learning was applied to sharpen performance in tool invocation, software engineering, and front-end development. Available under the MIT license, it can be used commercially and adapted for secondary development. The model balances competitive benchmark performance with superior efficiency, making it a practical choice for developers building cost-conscious AI agents that require reliable reasoning and coding capabilities.