NewsBite analysis · AI, Programming, Security, Startups · New Product, Software Release
MarkTechPost
Z.ai Ships GLM-5.3 Without Retraining the Base Model
Z.ai has introduced its latest model, GLM-5.3, which runs on the existing 743B base model of GLM-5.2, achieving remarkable increases in performance through post-training enhancements. Key benchmarks in coding and cybersecurity have seen substantial gains, with Terminal-Bench 3.0 scores dramatically improving from 4.6 to 28.3. The cybersecurity evaluation, CyberGym, exceeded expectations, reaching a score of 84.5%. The company plans to release model weights within the next two weeks after completing safety evaluations.
Act nowConfidence: HighImmediate horizon
Published 14 Aug 2026, 08:41 UTC2 min read
Visit original article ↗Lead image — MarkTechPost
Summary
TL;DR
- GLM-5.3 runs on the 743B base model of GLM-5.2.
- Coding performance improved significantly in Terminal-Bench 3.0.
- Cybersecurity metrics exceeded expectations at CyberGym.
- Weights for GLM-5.3 will be released in two weeks.
- Startups can adopt GLM-5.3 now via the Coding Plan or API.
Interesting facts
- GLM-5.3 achieved a 50% improvement over its predecessor on Z.ai Code Bench.
- The model demonstrates coalesced reasoning capabilities across exploitation tasks.
Coverage gaps