Zhipu AI releases GLM-5.3

Published: August 14, 2026 — THE DECODER

  • Same base as GLM-5.2; all gains from extended post-training
  • Claims strongest open-weights coding model; biggest jumps in agent tasks
  • Cybersecurity post-training produced multi-stage exploit-chain reasoning
  • Vendor: 2,436 vulns across 269 projects with Chinese security teams
  • Available via GLM Coding Plan; works with ZCode, Claude Code, OpenCode
  • Weights open in ~two weeks after security reviews