GLM-4.5 Air (EmberCloud)

GLM-4.5 Air is the compact counterpart in the GLM-4.5 family from Zhipu AI and Tsinghua University, designed to bring agentic, reasoning, and coding capabilities to a more accessible parameter budget. The companion technical report positions the family as open-source Mixture-of-Experts models, with GLM-4.5 itself described as carrying 355B total parameters and 32B activated parameters, while Air is released as the smaller sibling at 106B parameters intended for researchers and practitioners who need a lighter footprint. This positioning makes Air well suited for experimental and production scenarios where full-scale frontier models are too costly or too large to deploy, but where the ARC-style strengths of the family are still desirable.

The underlying training recipe for the family is described as multi-stage learning on 23T tokens, followed by post-training that combines expert model iteration with reinforcement learning, yielding a hybrid reasoning method that can switch between extended thinking and direct response modes. The flagship GLM-4.5 results reported in the same paper include 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified, with an overall third-place ranking among evaluated models and a second-place finish on agentic benchmarks, and Air is intended to inherit much of that behavior at a smaller scale. Open-weight availability through the zai-org repositories on Hugging Face, alongside the published technical report, makes the model a practical fit for teams building agent pipelines, code assistants, or reasoning-heavy tools that benefit from the family's hybrid reasoning design.