Antares: a 3B vulnerability-localization model approaches GPT-5.5 and beats open models 200x its size at $0.002 per task

arXiv·high signalXBlueskyLinkedInCopy link

Posted 2026-08-03 (arXiv 2608.02407), Antares is a family of 350M, 1B, and 3B agentic vulnerability-localization models built on IBM Granite bases, trained with supervised fine-tuning on cybersecurity reasoning plus repository-exploration data, then reinforcement learning from verifiable rewards over vulnerable repositories. Antares-3B approaches GPT-5.5 while outperforming open-weight models over 200x larger. The economics are the story for builders: a full 500-task evaluation sweep finishes in about 15 minutes on a single H100, amortizing to under 2 seconds and under $0.002 per task, which puts continuous local security scanning inside a CI budget instead of a frontier-API budget.