Z.ai / Zhipu AI Models

Z.ai publishes the GLM model family and its deployment resources. This directory retains the existing provider URL while distinguishing checkpoint names and serving routes.

GLM 5.3 and the earlier baseline

GLM 5.3 is an open-weight text release built on the same base as GLM 5.2. The publisher attributes the changes to post-training. Compare the versions on the same repository, tools and time budget instead of treating a reported benchmark gain as a universal improvement.

Serving settings affect behavior

The 5.3 card documents low, high and max reasoning effort, with max as the default when absent or invalid. It also documents chat-template controls such as clear_thinking. Record these settings alongside the checkpoint revision and tokenizer. A different template can change tool parsing and conversation behavior.

Licence and operational fit

GLM 5.3 has a named model licence. Downloadable weights should not be labelled Apache or MIT without reading that exact licence. The FP8 and BF16 checkpoints also have different deployment implications.

Start with tasks that expose failures: an issue with a regression test, a tool call with strict arguments and a request that lacks a required fact. Track accepted results, repeated ineffective actions and the work needed to repair an answer. The 5.2 versus 5.3 guide explains how to make the migration measurable.

Publisher resources

Current publisher documentation