Zhipu launches GLM-5.3 with strong coding claims, but the weights are not out yet
Originally announced by Z.ai . We link the primary source so you can read it for yourself.
Zhipu released GLM-5.3 on 14 August, and the reaction has centred on its coding performance, with the company describing it as the strongest open-weight coding model currently available. For anyone weighing it for local use, two things are worth being clear about before the excitement settles.
The weights are not available yet
GLM-5.3 launched through Zhipu’s hosted GLM Coding Plan and ZCode service, and it works with coding integrations including Claude Code and OpenCode. What it is not, at the time of writing, is downloadable. Zhipu says the open weights will follow in roughly two weeks, after a security review. Until then it is a paid hosted product, not something you can run on your own hardware, which is the part that matters most for a local AI audience.
The benchmark figures are the company’s own
The headline claim is a large jump in coding ability over GLM-5.2, and the numbers being shared are striking: Zhipu reports 28.3 on Terminal-Bench 3.0, up from 4.6 for the previous version, and improvements across agentic coding and cybersecurity evaluations, including scores it places alongside leading closed models on CyberGym. Every one of these figures comes from Zhipu’s own launch materials. No independent evaluations have appeared yet. That does not make them wrong, but claims of this size, particularly the ones putting an open model level with the frontier, are worth treating as a starting point until third parties test them.
One detail is genuinely interesting regardless. Zhipu says GLM-5.3 shares the same base as GLM-5.2, a roughly 744-billion-parameter mixture-of-experts model with about 40 billion active per token, and that the gains came from extended post-training rather than a new pretraining run. If the improvement holds up independently, it is a notable example of how much headroom post-training still has. The 1-million-token context window carries over from the previous version.
What we will do
When the weights are released we will add GLM-5.3 to the model catalogue and cover it in full, with the confirmed licence, the real hardware requirements, and any independent benchmark results by then available. It is worth flagging now that even once it is open, a 744-billion-parameter model is server-class: running it means a cluster or a high-memory server, not a consumer graphics card. GLM-5.2 was released under an MIT licence, so a permissive release is plausible, but we will confirm rather than assume. For now, treat GLM-5.3 as a promising hosted release whose open-weight version, and independent verification, are still to come.