Z.ai pushes GLM-5.2 open weights near frontier

Z.ai researcher Zixuan Li introduced GLM-5.2 on AI Engineer as an open-weights model positioned near frontier intelligence.

Li positioned GLM-5.2 between Claude Opus 4.7 and 4.8 on Terminal Bench 2.1 and other long-horizon tasks, highlighting its new "High" thinking level that boosts token efficiency.

Li emphasized that GLM-5.2 goes beyond coding tasks, showing strong gains on GDPval and math benchmarks plus an Artificial Analysis Intelligence Index score nearly matching frontier labs.

Zhipu AI, which operates under the Z.ai brand internationally, keeps the GLM name from its 2021 paper despite moving to different architectures.

Li defended the open weights approach for on-premise security and control, citing use cases like Harvey for domain fine tuning and co-design where customers can inspect the architecture and training recipes.

He wrapped up by releasing Z Code, a harness built for GLM-5.2 that also supports other frontier models with bring-your-own-key and Claude Code-style workflows.

That openness now raises alarms from Anthropic, which found through red teaming that GLM-5.3 exhibits Mythos-class autonomous cyber capabilities with weak safeguards against open weights misuse.

Teams considering local deployment gain auditability and cost control, but they also inherit the containment burden since an open weight cannot be recalled once downloaded.