Z.ai Releases GLM 5.3 — Coding +50%, Emergent Cyber Capabilities

Z.ai released GLM 5.3 with the same 743B parameters, but gains through RL post-training: +50% on Code Bench, discovery of 2,436 vulnerabilities in 269 open-source projects, API-breaking change — three thinking levels, cannot be disabled.

Compiled by Sergey KostenchukPublished 2026-08-14Updated 2026-09-18

2026-08-14 Research

🤖 Z.ai Releases GLM 5.3

A 743 billion parameter model, gains exclusively through RL post-training. +50% on Z.ai Code Bench, from 4.6 to 28.3 on Terminal Bench 3.0, 66.9 on DeepSWE. Unforeseen side effect: 2,436 vulnerabilities in 269 open-source projects.

🌍 Scaling laws for RL confirmed — without changing architecture, only post-training delivers breakthroughs. The model found vulnerabilities in Linux and WebKit, 1,097 of medium and high severity. Weights — via ~2 weeks after security review.

👤 API available with three thinking levels (low/high/max). Cannot disable thinking — breaking change for legacy applications.

Source: https://z.ai/blog/glm-5.3

Sources