GLM-4.5-Air is the streamlined sibling in Z.ai's GLM-4.5 family, designed to bring agentic AI capabilities within reach without flagship-scale compute. It uses a self-developed Mixture-of-Experts architecture with 106 billion total parameters but only 12 billion active at inference, a sparse design that keeps generation fast and inexpensive while still handling complex tasks. The model is purpose-built for intelligent agents, unifying reasoning, coding, and tool use inside a single system so that multi-step workflows can be coordinated end-to-end. Its defining design choice is hybrid reasoning: a "thinking mode" is available for multi-step planning and tool orchestration, while a "non-thinking mode" delivers immediate replies when depth is not required, letting applications pick the right balance of speed and deliberation for each call.

Within the same released lineup, GLM-4.5-Air is positioned as the efficiency-focused variant, with the flagship GLM-4.5 reaching 355 billion total parameters and both models evaluated across 12 industry-standard benchmarks. The Air version delivers a strong average score of 59.8 across that suite, trailing the flagship's 63.2 while consuming substantially fewer active parameters per token, which the team describes as achieving "competitive results while maintaining superior efficiency." Practical strengths show up clearly in agent benchmarks, where independent leaderboards report a 0.940 Tool Selection Quality score at a fraction of frontier-model cost, and in everyday responsiveness, with sub-second reply times in non-thinking mode. Available under the MIT license, the weights, base models, and FP8 variants were all open-sourced alongside the API release, making GLM-4.5-Air well suited to high-volume agent deployments, function-calling pipelines, and tool orchestration at scale.