Cantina and Yeta Labs release open-weight security research model apex-flash-1

Cantina Security and Yeta Labs released apex-flash-1, an open-weight safety research model, fine-tuned from Z.ai's GLM-5.3-Flash with GRPO reinforcement learning, published on Hugging Face under the MIT license. The model has 321.3B total parameters, with training data from 150 tasks built from 50 real vulnerability cases; it solved 40 of 60 tasks from 20 held-out vulnerability cases (66.7% pass@1), with a single-run cost of about $2.38, while Claude Opus 5 High solved 43 tasks (71.7%) at a cost of about $74.68. Deploying BF16 weights requires roughly 640 GB of GPU memory. The above evaluations and costs are internally reported benchmark results from the company, relayed by MarkTechPost; they should not be treated as independent evaluations or general real-world effectiveness.