Summary
Z.ai released GLM-5.3 on August 14, 2026, a coding-and-agent model built entirely through scaled post-training on its existing base rather than a new pretrain. It posts frontier open-weights coding scores (Terminal-Bench 3.0 jumped from 4.6 to 28.3) and an outsized cybersecurity gain (CyberGym 84.5%), and on launch day it privately reported a potentially serious vulnerability in Cursor. It is available now via Z.ai's API and GLM Coding Plan, with open weights promised in about two weeks after safety hardening.
What changed
Z.ai launched GLM-5.3 with major coding and cybersecurity gains derived from reinforcement post-training on the unchanged GLM-5.2 base model; on launch day the model privately disclosed a Cursor vulnerability that Cursor and Z.ai are jointly fixing.
Why it matters
GLM-5.3 pushes open-weights models to the frontier on long-horizon coding while showing that security-focused RL can yield autonomous vulnerability discovery strong enough to find real bugs in shipping developer tools, sharpening both the open-vs-closed model race and the case for AI-driven security review.
Evidence excerpt
Coding jumps most on the longest-horizon benchmarks, with Terminal-Bench 3.0 moving from 4.6 to 28.3; CyberGym reached 84.5%. The model found a potentially serious vulnerability in Cursor, disclosed privately, and the Cursor team is working with Z.ai on a fix.