Zhipu releases GLM-5.2 with a 1M-token context

Background

The benchmark figures in the release are given against GLM-5.1.^[1]

What happened

Zhipu released GLM-5.2 on 16 June 2026 under the MIT License.^[2][1] The context window was raised to 1M tokens, up from 200K.^[2][1]

Key details

Wikipedia gives the base model as 743 billion parameters, the same base later used by GLM-5.3.^[2] IndexShare reuses one sparse-attention indexer across every four layers, cutting per-token floating-point operations by 2.9 times at a 1M-token context.^[1]

On reasoning, GLM-5.2 scored 40.5 on HLE against 31 for GLM-5.1, and 54.7 against 52.3 with tools. It scored 91.2 on GPQA-Diamond against 86.2, and 99.2 on AIME 2026 against 95.3.^[1]

On coding, it scored 62.1 on SWE-bench Pro against 58.4, 81.0 on Terminal Bench 2.1 against 63.5, and 46.2 on DeepSWE against 18.^[1]

Impact

The longer context and lower per-token computing cost are the two architectural changes the release reports. Together they matter most for long agent tasks, where the model keeps a large amount of history in view.^[1]

Reading the benchmark figures

A benchmark score is only meaningful when two models are run on the same test under the same conditions. Vendors usually publish comparison tables that set their model against selected rivals on selected tests, so a number should be read together with the name of the test and the models it is compared with. A small gap between two scores does not by itself show a real difference in ability, and one test rarely captures everything a model can do.

Open weights and hosting

Open weights means the trained parameter files of a model are published so that others can download and run it. Open weights are not the same as open source, because whether a model counts as open source also depends on its licence, which decides what users may do with the weights. Hugging Face is a platform that hosts models and their weights for download, and several of the Zhipu models in this series were posted there.