GLM-4.5-Air is the release that established efficiency as a first-class goal for the GLM model family. It is a small-footprint model designed to deliver strong performance for every unit of compute it consumes, rather than to top a leaderboard — the tier you reach for when a workload is high-volume, latency-sensitive or running somewhere a flagship simply cannot be deployed.
That design goal has aged well. The family now offers a clear efficiency ladder — GLM-4.5-Air, then the 30B GLM-4.7-Flash, then the full-size releases, up to the 1M-context GLM-5.2 flagship — and all of it sits under the same MIT license with weights on the same public hosts. Choosing between them is a hardware and cost-of-compute decision, not a licensing one, which is a materially different position from choosing between hosted APIs where the tiers are also commercial boundaries.
No sign-up, no account, no credits. Type a prompt and get a streamed answer.
Response will stream here...
This playground runs a free open-source LLM via OpenRouter. It is not GLM-5.2 and does not use GLM weights — it's here so you can test a live model while you read the published GLM benchmark data below.
Frontier benchmarks reward capability at any cost, which is the opposite of how most production systems are actually constrained. If a request has to answer in under a second, or a pipeline has to process millions of documents, the operative question is not which model scores highest but which model clears your quality bar most cheaply. An efficiency-first GLM model changes the shape of what you can build: you can afford to call it on every item rather than sampling, run it inside a request path instead of a nightly batch, and keep the whole workload on your own hardware. The corollary is knowing when it is the wrong choice — long-horizon agentic work, repository-scale reasoning and hard multi-step problems belong on the flagship, and trying to force them onto a lightweight model produces expensive retries rather than savings.
The published comparison table covers GLM-5.2 and GLM-5.1. The column below is GLM-5.1's, shown only as a family reference point. No benchmark column has been published for GLM-4.5-Air, and none is invented here.
GLM-4.5-Air has no row in the published comparison set, so this page quotes none for it. The reference column exists to tell you what a full-size model from the following generation achieved, which is the honest way to frame a lightweight tier: you are trading measurable capability for a much smaller serving footprint, and the size of that trade is something you should measure on your own workload.
Comparison on HLE (Humanity's Last Exam), as published by the model authors.
HLE is the published evaluation that most punishes limited capacity. Claude Opus 4.8 leads at 49.8 on the full set, Qwen3.7-Max posts 41.4, GLM-5.2 40.5, DeepSeek-V4-Pro 37.7 and GLM-5.1 31.0. Even the strongest models answer fewer than half the questions, which tells you plainly that hard, open-ended reasoning is not where a lightweight model should be pointed — route those tasks to the flagship and keep Air for the high-volume work it was built for.
GLM-4.5-Air is the earliest model on this site's lineup and the origin of the family's efficiency line, later continued by the 30B GLM-4.7-Flash. Above it sit GLM-4.7, GLM-5, GLM-5-Turbo, GLM-5.1 and the GLM-5.2 flagship. A sensible architecture uses both ends: Air or Flash for volume, GLM-5.2 for the problems that justify the compute. Because every tier shares a licence and a weight host, that split costs you nothing contractually — you are running two checkpoints of the same open family rather than negotiating two separate vendor relationships, and you can move traffic between them whenever your measurements say you should.
The 1M-token context flagship. It is the open-source state of the art on coding and long-horizon evaluations, with explicit Non-Thinking, High and Max effort levels.
The lightweight 30B tier. Efficient and fast enough for local inference, and reported to outrank open models of a similar scale.
Read the GLM model benchmarks, then put a real model to work. The playground above is free and needs nothing from you; if you want a fuller AI toolkit, our partner's free tier starts here.