FrontierAugust 14, 2026via MarkTechPost
Z.ai Ships GLM-5.3 Without Retraining the Base Model: Better at Complex Coding and Long-Horizon Tasks
Why it matters
GLM-5.3 demonstrates that frontier capability gains no longer require expensive base-model retraining; scaled post-training on task-specific environments delivers outsized gains on reasoning and coding—a shift in where labs are investing.
Key signals
- GLM-5.3 released August 14, 2026
- Reuses 743B GLM-5.2 base unchanged
- Terminal-Bench 3.0: 4.6 → 28.3 (+515%)
- DeepSWE v1.1: 46.2 → 66.9 (+45%)
- CyberGym: 84.5%
- ExploitBench: doubled to 54.4%
- Gains from scaled post-training only (longer training, more environments, more task types)
- Weights available in ~2 weeks
- Cybersecurity gains exceeded Z.ai's reported plans
The hook
Z.ai just shipped a 743B model that punches way harder on coding and long-horizon tasks—without touching the base weights. Post-training scaling alone moved Terminal-Bench from 4.6 to 28.3.
Z.ai released GLM-5.3 on August 14, 2026. The model reuses the 743B GLM-5.2 base unchanged. Every reported gain comes from scaled post-training: more long-horizon task environments, more environment types, longer training. Terminal-Bench 3.0 moves from 4.6 to 28.3, and DeepSWE v1.1 from 46.2 to 66.9…