Z.ai Launches GLM-5.3
Published: August 14, 2026 — Unite.AI
Vendor-reported benchmarks:
- CyberGym: 84.5% (vs GLM-5.2 77.2%)
- ExploitBench: 54.4% (vs 24.4%); Mythos 5 at 78.0%
- ExploitGym: 105 tasks/2h, 130/6h (vs 29/39 for 5.2); Mythos 5 at 181/247
- Open weights delayed ~2 weeks for safety hardening of cyber capability
- Notes need to independently replicate company harness scores