Z.ai's most capable model (released Aug 14, 2026). Same base model as GLM-5.2 — **every gain comes from scaled post-training** on long-horizon environments. 50% improvement over GLM-5.2 on internal Z.ai Code Bench with fewer output tokens. **Domain-leading strengths: cybersecurity (CyberGym SOTA), vulnerability exploitation (ExploitBench 2.2× GLM-5.2), terminal/agentic coding (Terminal-Bench 3.0 open-source SOTA), long-horizon autonomous engineering (SWE-Marathon +119%, AutomationBench +84%), and professional knowledge work (GDPval-AA v2 1769 Elo).** Found 2,436 real-world vulnerabilities across 269 projects (1,097 medium-to-high severity). Always-on reasoning with configurable effort (`low`/`high`/`max`; default `max`). Token-efficient: at High effort, 31.4% on Z.ai Code Bench at ~50K tokens surpasses Opus 4.8's 29.5% at 120K. MoE 744B/40B-active. Weights to be released ~2 weeks after launch under open-source license. **Caveats: trails GPT-5.6 Sol on Terminal-Bench 3.0 (28.3 vs 34.6), DeepSWE v1.1 (66.9 vs 72.7), and HLE (62.5 vs 64.5); trails Fable 5 on ExploitBench (54.4% vs 78.0%) and ExploitGym (105/130 vs 181/247 tasks); all benchmarks vendor-reported with limited independent verification at launch; reasoning cannot be disabled (must use `low` at minimum).**
Architecture Type: Mixture of Experts (MoE)
Sparse Mixture-of-Experts routing: activates only a subset of parameter experts per token, delivering frontier-level intelligence with exceptional efficiency.
Recommended Workloads & Primary Use Cases
Everyday reasoning
Drafting
Fast problem solving
ARMES Zero Data Retention
Calls to GLM-5.3 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Z.ai.
