Cisco's Foundation AI group released two very small open-weight models on July 21 — Antares-350M and Antares-1B — built for one narrow job: finding where a known vulnerability lives inside a codebase. A 3B version was announced but deliberately not released.
Localization, not discovery
The distinction is the whole product. Antares does not hunt for unknown bugs; given a vulnerability description, it works agentically — searching for code patterns, reading candidate files, folding in evidence and backtracking when a path fails — to pinpoint the affected code. Cisco calls it "one of the hardest, most time-consuming and expensive problems in security."
The benchmark numbers
On Cisco's own VLoc Bench, 500 tasks scored on File F1: GPT-5.5 leads at 0.229 and 0.221, with Antares-3B at 0.223 and Antares-1B at 0.209. Below them sit GLM-5.2, a 753-billion-parameter model, at 0.186, and Gemini 3 Pro at 0.152. A one-billion-parameter model beating a 753-billion one is the headline result.
And the cost gap
A full 500-task evaluation costs about $0.82 with Antares, against roughly $141 for GPT-5.5 and $12.50 for GLM-5.2 — about 172x cheaper than the frontier model it nearly matches. It completes in about 13 minutes on a single H100 with 16 parallel workers, versus roughly five hours for frontier models. Antares is built on IBM Granite 4.0 1B, licensed Apache 2.0, and runs locally so source code never leaves the organisation.
Open, but gated
Access is vetted. "We are completely gating the 3B model to make sure that we responsibly release it to communities that need it," said Amin Karbasi, Cisco's VP and chief AI scientist. The benchmark and cost figures are Cisco's own, on Cisco's own benchmark, with no independent replication yet.