GLM models

The family

GLM is Zhipu AI's open-weight line, released under a modified MIT licence with an explicit patent grant. It sits at the other end of the catalogue from the rest of what we serve: trillion-scale mixture-of-experts models aimed squarely at the frontier, published as weights anyone can run rather than as an API.

One GLM model is on offer here, and it is the largest model we serve by a wide margin.

What we serve

Size and shape

744B total parameters with roughly 40B active per token. The checkpoint is 433 GB on disk even at NVFP4, which is more than any single card holds: it serves across four NVLinked B200s as one tensor-parallel worker. That is why it is the most expensive model on the rate card.

Input

Text only. There is no image, audio or video encoder in these weights - anything with media in it goes to Qwen or Gemma.

Output

Text, and it thinks by default: reasoning is this checkpoint's standard mode and the trace bills as output. A job can put a reasoning budget on it, which is worth doing at this price.

What it is served from

NVIDIA's own NVFP4 quantization of the published weights, not one we produced.

Benchmark scores

Published figures, each from the source named beside it. The agentic rows come from a vendor harness and have not been independently reproduced, so treat their margins as directional.

For scale: an index of 51 puts it within a few points of the closed frontier models and 14 points above the best-scoring 27B checkpoint we serve. That gap is real, and so is the price gap - glm-5.2 costs a little under three times what qwen3.6-27b does on a typical mix of input and output tokens, before you count the thinking trace it emits by default.

When to pick it

Work the smaller models get wrong

Hard reasoning, long multi-step derivations, code that has to be correct rather than plausible. This is the model to reach for when a 27B run came back not good enough, and it is worth measuring that rather than assuming it.

A quality ceiling for a sample

Running a few hundred rows through glm-5.2 and the rest through a cheaper model is a good way to find out how much the price difference actually buys you on your own data.

Not for

Bulk classification and extraction, or anything with an image, audio or video in it. The first is a waste of the price, the second it cannot do at all.

Pass glm-5.2 exactly as written when you submit a job. Its rates are on the pricing page, and because every job is estimated before it runs, you see what a GLM run costs before you approve it.