GLM-4.5-Air

GLM-4.5-Air was published the same day as the flagship it shrinks, in late July 2025, and shares its entire training pipeline: 15 trillion tokens of general pretraining, fine-tuning on code, reasoning and agent tasks, then reinforcement learning. What differs is scale. Z.ai reports 106 billion total parameters with 12 billion active per token, against 355 billion and 32 billion for GLM-4.5; the safetensors index on Hugging Face totals 110.5 billion. Everything a reader would use to tell two models apart stays the same here. The context window is 128,000 tokens. Hybrid reasoning works identically — Thinking Mode for tool chains and hard problems, Non-Thinking Mode for instant replies. The licence is MIT, so the weights can be downloaded, modified and deployed commercially without asking anyone. The interesting part is the price sheet. Air costs $0.20 per million input tokens and $1.10 per million output — a third of the flagship's rate. But Z.ai also sells AirX, the same model on faster infrastructure, at $1.10 and $4.50: roughly double what the three-times-larger GLM-4.5 costs in its standard tier. At Z.ai, speed is a more expensive commodity than size, and a reader picking a variant should compare the whole grid rather than assume the smaller model is always the cheaper choice. On Hugging Face, Air is downloaded more often than the flagship — 172,000 pulls against 115,000 — which is the usual pattern for open weights: the model people can actually fit on their own hardware wins on volume.

▸Videos

No videos yet.

▸Reviews