GLM-4.7-Flash is the 30-billion-parameter member of the GLM model family, built for the case where the flagship simply will not fit. Thirty billion parameters is the size class that runs on a single high-memory accelerator, or quantised on consumer hardware, which makes it the natural choice for local development, on-premise inference and anything where per-request latency matters more than the last few benchmark points.
The model authors position it as efficient and high-performance, and report that it outranks open-source models of a similar scale. That claim is the relevant one for this tier: nobody deploying a 30B model expects it to beat a frontier system, but they very much care whether it beats the other 30B options they could have run instead.
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.
Parameter count decides where a model can physically run, and that constraint drives more architecture decisions than raw capability does. A 30B model can be served on hardware a small team already owns, keeps prompts inside your own network, and answers fast enough to sit in an interactive loop rather than a batch queue. It is also the size at which fine-tuning stops being a research project — you can adapt a 30B GLM model to a narrow domain and still evaluate the result in an afternoon. The trade is real: fewer parameters means less headroom on hard, long-horizon reasoning, which is exactly the workload the 1M-context flagship was built for. The right way to use the family is to route work by difficulty, keeping cheap, high-volume tasks on Flash and reserving the flagship for problems that genuinely need it.
The published comparison table covers GLM-5.2 and GLM-5.1. The column below is GLM-5.1's, shown only as a reference point for the family. No benchmark column has been published for GLM-4.7-Flash, and none is estimated here.
There is no published GLM-4.7-Flash row, and inferring one from a much larger model would be misleading in either direction. What the reference column is good for is calibration: it shows the spread the published GLM models occupy, so you know what a full-size model of the previous generation achieved before you decide whether a 30B model is likely to be adequate for your task.
Comparison on DeepSWE, as published by the model authors.
DeepSWE is where model scale shows up most brutally. GPT-5.5 leads at 70.0 and Claude Opus 4.8 posts 58.0, GLM-5.2 reaches 46.2, and the previous open generation clusters far below — GLM-5.1 and Qwen3.7-Max at 18.0, DeepSeek-V4-Pro at 8.0. If your workload looks like DeepSWE, a lightweight model is the wrong tool no matter how efficient it is; if it looks like classification, extraction or routine code edits, Flash is likely all you need.
GLM-4.7-Flash is the lightweight tier of the GLM-4.7 generation, sitting alongside the full-size GLM-4.7 and above the older GLM-4.5-Air in the family's efficiency line. Everything in that line carries the same MIT license and the same weight hosts, so moving between tiers is a hardware decision, not a legal one.
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 original efficiency milestone of the family: a small-footprint GLM model with strong performance for every unit of compute it consumes.
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.