Antares scans 500 vulnerability tasks in 15 minutes

Antares-3B completed a 500-task vulnerability-localization sweep in about 15 minutes on one H100, averaging under two seconds and less than $0.002 per task, according to the 57-page Antares technical report. The IBM Granite-based family spans 350M, 1B, and 3B parameters; its authors say the 3B model approaches GPT-5.5 and beats open-weight models more than 200 times larger, but those are paper results awaiting independent reproduction. Security teams tracking AI’s expanding role in vulnerability discovery should test compact local triage before routing every repository scan to a frontier API.