- Zhipu says GLM-5.3 uses the same base model as GLM-5.2 and gains its reported improvement from post-training alone. [1] - The company says security teams found 2,436 vulnerabilities across 269 projects, but the result is a Zhipu claim rather than an independent evaluation. [1][2] - Weights are scheduled for release in about two weeks after security reviews, so independent benchmark and tester results are not yet available.
[1] Zhipu AI released GLM-5.3 on August 14, 2026, through its GLM Coding Plan and claimed that the model is the strongest open-weights coding system currently available. The company says GLM-5.3 shares the same base as GLM-5.2 and achieved its gains through extended post-training rather than a new pretraining run. [1]
Zhipu also positioned the release as a cybersecurity model. It says the system helped security teams identify 2,436 vulnerabilities across 269 open-source projects, including flaws dating back as far as 45 years. Z.ai’s public vulnerability ledger currently lists the same total, with 53 disclosures published and 2,383 entries not yet public. [2]
Evidence is limited until the weights arrive #
GLM-5.3 currently works with Zhipu’s ZCode, Claude Code and OpenCode integrations, but the model weights are not yet downloadable. Zhipu said it plans to release them in approximately two weeks after security reviews. [1]
That timing limits what can be established today. The company’s claim of a 50% improvement and its “strongest open-weights coding model” description are based on Zhipu’s own evaluations; the reviewed release materials do not yet provide an independently reproduced benchmark set for GLM-5.3. The weights license, model card and deployment behavior should be checked when the files are published. [1]
Further sources #
The stories that matter, in one email. Free — unsubscribe anytime.