GLM-4.5-Air is the leaner sibling in the GLM-4.5 family, shaped as a Mixture-of-Experts model with 106 billion total parameters and 12 billion active parameters. Z.ai released it on July 28, 2025, alongside the larger GLM-4.5, framing the pair as flagship foundation models aimed squarely at intelligent-agent applications. Rather than separating chat, reasoning, and tool use into different checkpoints, the design unifies reasoning, coding, and agentic capabilities into a single model that can toggle between a thinking mode for complex multi-step work and a non-thinking mode for instant replies, making it practical to deploy in agent loops where latency and depth trade off turn by turn.
Beyond the Z.ai chat surface and the documented API, GLM-4.5-Air is distributed as open weights on Hugging Face under the zai-org GLM-4.5 collection, with a companion GitHub repository and an arXiv technical report (2508.06471) describing the architecture. Third-party hosts such as Fireworks AI offer on-demand deployment on dedicated GPUs under the model path accounts/fireworks/models/glm-4p5-air, which gives teams a route to serve the model without standing up their own inference stack. For builders assembling agentic systems that need coding and tool use in one backbone, the Air variant offers a noticeably smaller active-parameter footprint than its 355B/32B sibling while keeping the same hybrid reasoning design.
