Cisco’s Foundation AI team introduced Antares, a family of security small language models purpose-built for vulnerability localization: pinpointing which files in a repository actually contain a known weakness. The open-weight Antares-350M and Antares-1B work like a human investigator with a terminal (search, read candidate files, backtrack when a path goes nowhere). They outperform a dozen larger closed- and open-weight models on Cisco’s new 500-task Vulnerability Localization Benchmark, while being compact enough to run locally so sensitive code never leaves the building. According to the technical report, the models start from IBM Granite checkpoints, are trained with supervised fine-tuning followed by reinforcement learning on verifiable file-level rewards, and the not-yet-released Antares-3B reaches a File F1 of 0.223, close to GPT-5.5’s 0.229. Near-frontier triage without the frontier invoice; the files it flags, however, still fix themselves at the usual rate of never.
