Can an Open Model Do Security Research? Cantina’s apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks

Cantina Security and Yeta Labs have put out apex-flash-1, an open-weights model tuned for vulnerability research rather than generic chat. It’s built by fine-tuning Z.ai’s GLM-5.3-Flash with reinforcement learning, and the weights are on Hugging Face under the MIT license. That makes it unusually practical for a security model: it’s meant to be run, not just admired from a slide deck.

The model is huge. Cantina says the safetensors metadata puts it at 321.3 billion total parameters, while the GLM-5.3-Flash base is a mixture-of-experts model with 18 billion active parameters. Training used GRPO, a rank-256 LoRA, and selective full-parameter training. The data itself came from 50 real vulnerability cases turned into 150 tasks, with each case split into guided whitebox, focused whitebox, and focused blackbox variants.

The mix of bugs says a lot about what the model was taught to care about. Authorization, identity and scope issues account for 72% of the cases. Accounting and numerical precision make up another 18%. The rest cover time validation, business rules and SSRF. Cantina says the reinforcement learning rollouts ran inside the Codex agent harness on production-like software and protocol environments.

On its internal benchmark, apex-flash-1 solved 40 of 60 held-out tasks, which is 66.7% pass@1. Cantina says that run cost about $2.38. The base GLM-5.3-Flash model solved 36 of 60 for about $4.56, while Claude Opus 5 High solved 43 of 60 for about $74.68. So Opus found three more tasks, but at roughly 31 times the cost per run. On Cantina’s math, that’s about $0.06 per solved task for apex-flash-1 versus $1.74 for Opus.

The catch is obvious. BF16 deployment needs about 640 GB of GPU memory, so this is not exactly a laptop hobbyist’s afternoon project. But Cantina’s pitch is aimed at defenders who want a capable worker model they can keep local and under control, rather than handing security work to a black-box service.