Z.ai's day-old GLM-5.2 successor, tuned for coding and cyber defense; weights gated at launch.
GLM-5.3 is Z.ai's successor to GLM-5.2, launched August 14, 2026 under the tagline "Built to Code. Ready for Cyber Defense." It is post-trained on a roughly 743-billion-parameter base and carries the family's 1M-token context, continuing GLM's focus on long-horizon, code-heavy agentic work — now with an explicit cyber-defense emphasis.
Two things set this release apart from GLM-5.2, and both counsel caution. Unlike 5.2's immediate MIT open-weight drop, GLM-5.3's open weights and API access are staged behind a safety review rather than shipped on day one — so it is not yet freely downloadable. And because it is barely a day old, benchmark coverage is thin and vendor-reported; no independent index exists yet. Vision, the community's most-requested addition, is not part of this launch — it remains text-focused.
For teams already invested in the GLM coding stack, 5.3 is the version to watch, and the cyber-defense positioning is a distinctive bet. But until the weights actually ship and independent numbers land, GLM-5.2 remains the concrete open-weight reference in this lab's line, and DeepSeek V4-Pro is still the independently verified open-tier benchmark.
FIELD NOTE — Launched Aug 14; open weights/API staged behind safety review. Benchmarks thin, vendor-only.
VISIT OFFICIAL SITE →