Can an Open Model Do Security Research? Cantina’s apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks
40 of 60. That's how many vulnerability-discovery tasks Cantina's open-weight apex-flash-1 solved — a reinforcement-learned model built specifically for security research, now available on Hugging Face under MIT.
Why it matters
An open-weights model fine-tuned for vulnerability research hits a meaningful capability threshold and ships deployable weights. For security teams and red teams, this trades frontier-model API access for local control and specialization; the catch is the 640 GB GPU memory requirement and unproven real-world deployment.
The key facts
12 to knowapex-flash-1 is a RL fine-tune of Z.ai's GLM-5.3-Flash
Solves 40 of 60 held-out vulnerability discovery tasks
Released on Hugging Face under MIT license
Deployable on vLLM, SGLang, or Transformers
BF16 inference requires ~640 GB GPU memory
Developed by Cantina Security with Yeta Labs
Model: apex-flash-1, RL fine-tune of Z.ai's GLM-5.3-Flash
Release: open-weights, MIT license, Hugging Face
Capability: solves 40 of 60 held-out vulnerability discovery tasks
Deployment: vLLM, SGLang, or Transformers; BF16 requires ~640 GB GPU memory
Developer: Cantina Security with Yeta Labs
Domain: vulnerability research and security
The story so far
Earlier coverage of this storyline
- Aikido Security Releases Altar-1: An Open-Weight Security Model Pruned From GLM-5.3 to 328 GBMarkTechPost
- 2026 in LLMs (so far)Simon Willison
- This story
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Cantina Security, with Yeta Labs, has released apex-flash-1, an open-weights model trained specifically for vulnerability research. It is a reinforcement learning fine-tune of Z.ai’s GLM-5.3-Flash, released on Hugging Face under the MIT license. Is it deployable? Yes, the MIT weights serve on vLLM,…