apex-flash-1

apex-flash-1 is an open-weight model for security research published by Cantina Security, developed with Yeta, on Hugging Face on 30 September 2026 under the MIT licence. It is a reinforcement-learning post-train of the GLM-5.3-Flash mixture-of-experts model, reported at about 321 billion parameters in total with 18 billion active, and its configuration allows a context of about one million tokens. Cantina describes it as built for focused investigations: reading code, using tools, pursuing a potential exploit and verifying its effect against a running target in an isolated environment. Its model card reports results on 60 tasks drawn from 20 held-out vulnerability cases and compares cost per run with larger closed models; those figures are Cantina’s own, so no score is stored here. MarkTechPost covered the release on 4 October. For security teams, an open model of this kind can be run on their own infrastructure for code review and penetration testing, but like any offensive-security tool it should be used only against systems they are authorised to test.

Not yet independently verified: The date is the Hugging Face repository’s creation time (30 September 2026); MarkTechPost covered it on 4 October. The benchmark results and cost comparisons are Cantina’s own, on its internal held-out set, so no score is stored. We will update this when it can be confirmed, and remove this note.