GLM-5.3
Coding and security-focused post-train of the GLM-5.2 base, with DeepSWE up from 46.2 to 66.9, CyberGym at 84.5 and weights under a new custom licence.
Zhipu says internal coding capability is 50% better than GLM-5.2 and that the model found 2,436 vulnerabilities across 269 projects. It ran its 'most extensive risk review' before releasing weights, and dropped MIT for a GLM-5.3 licence.
- Date
- Friday, 14 August 2026
- Lab
- Zhipu AI / Z.ai
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| DeepSWE v1.1 | 66.9 from 46.2 for GLM-5.2 | company |
| CyberGym | 84.5% highest on benchmark per Z.ai | company |
| ExploitBench | 54.4% from 24.4% | company |
| Vulnerabilities found | 2,436 in 269 projects as reported by The Decoder | company |
Subscriber launch 2026-08-14 per The Decoder and Apidog; Z.ai release notes list 2026-08-18. HF repo created 2026-08-25 with custom licence. The 2,436-vulnerability figure appears in The Decoder only, not in the Apidog write-up.
Sources
- the-decoder.com/zhipu-ai-releases-glm-5-3-claims-its-the-strongest-open-weights-coding-mod
- docs.z.ai/guides/llm/glm-5.3
- huggingface.co/zai-org/GLM-5.3
- apidog.com/blog/what-is-glm-5-3/
This record was checked against its sources on 6 October 2026. How we check