Cantina releases apex-flash-1, an open-weights vulnerability research model

Cantina Security and Yeta Labs released apex-flash-1, an MIT-licensed open-weights MoE model with 321.3B total and 18B active parameters, fine-tuned from GLM-5.3-Flash. On an internal 60-task benchmark it solved 40 (66.7% pass@1) versus 43 for Claude Opus 5 High, at $2.38 per run against $74.68.

• 321.3B total parameters, 18B active, MIT license
• Solved 40 of 60 tasks, 66.7% pass@1
• Claude Opus 5 High solved 43 but cost 31x more
• GRPO training on 150 tasks from 50 real vulnerability cases