Z.ai's new GLM 5.3 model will become available on the Vercel AI Gateway to assist developers with cybersecurity and coding tasks. This deployment focuses on the model's specialized performance in the DeepsecBench benchmark and its operational efficiency. The integration allow developers to execute security evaluations more frequently due to the lower cost compared to proprietary tools with similar output quality.
The model utilizes a 743B base model shared with GLM 5.2 but improves performance through post-training in simulated engineering environments. These updates boosted Terminal-Bench scores from 4.6 to 28.3 and raised DeepSWE from 46.2 to 66.9. On the CyberGym benchmark, GLM 5.3 achieved an 84.5% score, surpassing the result for Anthropic's Mythos 5.
Key sources
- SOURCE@vercel_dev“1/3 of the cost of some proprietary models with similar scores”x.com
- SOURCE@rauchg“means you can run https://t.co/KwrlahlCiK at least 3× more often!”x.com
- SUPPORT@wesroth“Terminal-Bench jumped from 4.6 → 28.3 and DeepSWE from 46.2 → 66.9”x.com
- SUPPORT@maxforai“GLM 5.3 已经站到了开源模型的第一梯队”x.com