Cantina Security releases apex-flash-1, a 321.3B open model for vulnerability research
Cantina Security, together with Yeta Labs, announced apex‑flash‑1, an open‑weights model built on Z.ai’s GLM‑5.3‑Flash and released on Hugging Face under an MIT license. The 321.3 B‑parameter model was fine‑tuned with GRPO, using a rank‑256 LoRA plus selective full‑parameter training on 150 tasks derived from 50 real vulnerability cases.
Key points
- Cantina Security and Yeta Labs released apex‑flash‑1, a 321.3B open‑weights security model under MIT license.
- In a 60‑task benchmark, apex‑flash‑1 solved 40 tasks (66.7% pass@1) at about $2.38, versus Claude Opus 5 High’s 71.7% at $74.68.
- The model runs on vLLM, SGLang or Transformers but needs roughly 640 GB GPU memory for BF16 inference.
In Cantina’s internal benchmark of 60 held‑out tasks, apex‑flash‑1 solved 40 of them, achieving a 66.7 % pass@1 rate at an estimated cost of $2.38 per run. By comparison, Claude Opus 5 High solved 43 tasks (71.7 % pass@1) but cost $74.68, roughly $0.06 per solved task for apex‑flash‑1 versus $1.74 for Opus. The model runs on vLLM, SGLang or Transformers, but BF16 inference requires about 640 GB of GPU memory. Cantina positions apex‑flash‑1 as a worker model that can be orchestrated by larger systems, aiming to give defenders a locally controllable security‑focused AI.
Model page: apex-flash-1 →
Can an Open Model Do Security Research? Cantina’s apex-flash-1 Solves 40 of 60 Held-Out Bug Tasks
MarkTechPost · 5 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Michal Sutter. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Aleph Alpha releases Kolibri-1 MoE model with 78B parameters and 1M token context · 4 src
- Anthropic releases Claude Sonnet 5.5 with 30% speed and cost gains · 4 src
- Microsoft launches MAI-Transcribe-2-Streaming real-time transcription AI for $0.54 per hour · 2 src
- OpenAI, Google, and Anthropic launch new AI models in late September · 2 src
- Anthropic asks users to share voice data for AI model training · 2 src
Comments
via GitHub Discussions