Cisco’s tiny Antares models find software bugs bigger models miss
Cisco released Antares-350M and Antares-1B, open-weight models that localize known vulnerabilities in codebases—matching models many times their size at a fraction of the cost.
In Brief
- Cisco released Antares-350M and Antares-1B, open-weight models built to pinpoint where known vulnerabilities live in a codebase.
- The 1B model hits 0.209 File F1 on the VLoc Bench, close to GPT-5.5 (0.229) and above a 753B open model (0.186).
- They run locally, keeping sensitive source code off third-party clouds—but a 3B version stays gated.
Cisco just made a quiet but useful argument: you don’t need a frontier model to find a security hole. Its Foundation AI group released Antares, a family of small language models purpose-built for one narrow, expensive job—vulnerability localization.
Give Antares a vulnerability description and a repository, and it finds the files that contain the flaw. Cisco shipped two open-weight versions, Antares-350M and Antares-1B, both under Apache 2.0, alongside a 500-task benchmark called VLoc Bench. A larger 3B model exists but is being held back from public release.
The point isn’t a new state of the art. It’s efficiency. As Cisco’s own blog puts it, the models are compact enough to run locally, “heading off the need to send sensitive codebases to the cloud”—a real concern when the thing you’re scanning is proprietary source.
What vulnerability localization actually solves
Software security depends on connecting public vulnerability knowledge—advisories, CVEs, CWE entries—to internal code that is large, modular, and dependency-rich. Today that first triage step is where cost concentrates: developers search unfamiliar code, follow naming conventions, and inspect call paths by hand.
On VLoc Bench, Antares-1B reaches 0.209 File F1. For context, The Register reports GPT-5.5 scores 0.229 and a 753B open-weight model scores just 0.186—so a 1-billion-parameter model edges out one roughly 750 times its size on this task.
Cisco VP and chief AI scientist Amin Karbasi framed the stakes bluntly: “The impact of vulnerabilities in your codebase is huge, but it might be only a single file or a few lines of code in a million lines of code.” The model’s job is to find that single file fast.
Why small open models beat sending code to the cloud
Antares does not replace the security toolchain. Cisco is explicit that teams still need dependency scanning, secret scanning, dynamic testing, container checks, threat modeling, and expert review. What changes is the first pass: a cheap, local model can narrow a million-line repo to a handful of candidate files before a human or a bigger model engages.
That local-first design matters because the alternative—shipping source to a hosted API—is a non-starter for many enterprises. Axios notes the models give companies a cheaper way to repeatedly search codebases without exposing them to an outside provider.
For defenders, the takeaway is that capability isn’t only at the top of the scale. As open-weight models improve on narrow security tasks, the economics of continuous code scanning shift from “too expensive to run often” to “cheap enough to run always.”
FAQ
Are Cisco’s Antares models free to use?
The 350M and 1B versions are open-weight under Apache 2.0, available on Hugging Face to verified defenders; a 3B model is held back.
How good are they?
On the 500-task VLoc Bench, Antares-1B reaches 0.209 File F1, outperforming models hundreds of times larger including GLM-5.2 and Gemini 3 Pro.
Do they replace security tooling?
No. Cisco says they handle first-triage localization only; dependency scanning, secret scanning, and expert review still matter.
