Z.ai released GLM-5.3, a 753-billion parameter model that improves coding and agentic capabilities through fine-tuning rather than training from scratch. The update ties the open-weights leader Kimi K3 on Artificial Analysis’ intelligence index and achieves top scores on cybersecurity benchmarks.

  • Architecture: Mixture-of-experts transformer with 40 billion active parameters per token, supporting up to 1 million input tokens.
  • Training: Utilizes single-rollout asynchronous optimization and compaction of long-running task segments to enhance learning from extended environments.
  • Cybersecurity: Achieved 84.5 percent on the CyberGym benchmark and 54.4 percent on ExploitBench, outperforming GLM-5.2 significantly.
  • Intelligence Index: Scored 60 points at max reasoning, matching Kimi K3 but trailing proprietary leaders like Claude Opus 5.

The model is available via Z.ai’s API and coding plans, with weights expected to release approximately two weeks after launch.