AIBRIEF
On a single Nvidia H100 AI accelerator chip, the small 350 million and 1 billion parameter versions of Antares — Cisco Foundation AI's new open weight (downloadable model files, not training data or full source code) model family — scan 290 real
Cisco Foundation AI released Antares, a 350M/1B/3B family of decoder-only transformers built to localize known CVEs inside real repositories. The 350M and 1B versions are open-weight on Hugging Face under Apache 2.0; the 3B stays gated to vetted communities.
Cisco re-fine-tuned IBM Granite 4.0 checkpoints and shipped the model as a read-only terminal agent: 15-call budget per task, Docker-sandboxed, networking off, output capped at 2,000 characters. The agent returns a ranked list of files suspected to carry a given CWE.
On Cisco's VLoc Bench (500 tasks, 290 repos, 147 CWEs), Antares-3B scores 0.223 File F1 against GPT-5.5's 0.229. Antares-1B hits 0.209 with the leaderboard's highest recall at 0.224. Against static analyzers on the same bench, Semgrep lands at 0.086, CodeQL at 0.023, Horusec at 0.020; Antares-350M at 0.135 already beats all three.
On a single H100, Cisco chief AI scientist Amin Karbasi told The Register the 500-repo sweep runs under $1 versus $100–$150 for frontier models, finishing in 11 to 15 minutes instead of roughly five hours.
The benchmark numbers come from Cisco and have not been independently reproduced. Stanford's Amin Saberi calls always-on scanning feasible at this cost; NUS's Reza Shokri warns the same agent helps attackers find bugs faster. Weights stay gated on Hugging Face, so the local-only benefit is real for vetted users, not universal.