Cisco has released two open-weight small language models, Antares-350M and Antares-1B, designed specifically to hunt for known vulnerabilities in existing codebases, positioning them as a leaner alternative to using large frontier AI systems from Google and OpenAI for security scanning. The models are available on Hugging Face but only to vetted users, including academic, nonprofit and smaller security teams, and are designed to run locally so proprietary source code never has to leave an organisation's own machines.

Cisco says both models perform as well as or better than much larger rivals on a new benchmark measuring vulnerability detection efficiency, with Antares-1B reportedly outperforming Google's Gemini 3 Pro and matching Z.ai's GLM-5.2, while an unreleased Antares-3B beats GLM-5.2 and OpenAI's GPT-5.5. The company claims Antares can scan 500 repositories in around 15 minutes for under $1, compared with roughly five hours and $100-$150 for frontier models, crediting this to training the models as targeted "investigator" search tools rather than general-purpose chatbots.