zai-org/GLM-4.5-Air

GLM-4.5-Air is the lighter member of the GLM-4.5 family of foundation models aimed at intelligent agents. It uses a Mixture-of-Experts style compact design with 106 billion total parameters and 12 billion active parameters, deliberately trading some of the flagship GLM-4.5's capacity for lower compute cost. The series unifies reasoning, coding, and agent-oriented capabilities inside a single architecture, so the same model can drive tool-using assistants as well as code generation workloads. The base, hybrid reasoning, and FP8 variants are released under an MIT license, which keeps the door open for commercial and secondary development without proprietary lock-in.

What distinguishes GLM-4.5-Air from a typical dense chat model is its hybrid reasoning behavior. At inference time it can switch between a non-thinking mode that produces fast, immediate responses and a thinking mode that allocates extra deliberation for complex reasoning and multi-step tool use, letting developers tune latency against answer quality per query. Across the family's 12-benchmark evaluation, GLM-4.5-Air posts a score of 59.8, landing close behind the flagship GLM-4.5 at 63.8 and ahead of many open-source peers, while its smaller active footprint makes it a practical fit for cost-sensitive deployments that still need strong agent and coding performance.