Cisco launches Antares SLMs to localize code vulnerabilities

Cisco Foundation AI announced Antares, a pair of open-weight small language models called Antares-350M and Antares-1B. The models and their parameter weights are available on Hugging Face.

Antares is designed to help security teams find where a reported vulnerability appears in a codebase. The system produces a ranked list of source files likely to contain the issue to help triage work; Cisco described the models as tools for initial investigation rather than final vulnerability determinations.

The models follow a search-style process that starts from a vulnerability description, searches for matching code patterns, reads candidate files, updates evidence, shifts focus if a path is unhelpful, and narrows toward files most likely to contain the problem. Outputs are ordered by likelihood to help analysts prioritize reviews.

Cisco released a 500-entry Vulnerability Localization Benchmark to measure performance. The benchmark requires models to scan unfamiliar repositories and recognize vulnerability patterns associated with specific Common Weakness Enumeration categories.

In tests run by Cisco, Antares completed the benchmark faster and at lower cost than a leading closed model and a leading open-weight general model. Cisco reported Antares was 172 times cheaper than the closed model and 15.2 times cheaper than the open-weight general model it tested against.

Antares is open-weight, meaning the models’ parameter weights are published so organizations can run and control the models locally. Cisco noted that open-weight does not automatically imply an open-source license covering redistribution; it refers specifically to access to the weights. Small language models like Antares use fewer parameters than large foundation models, reducing compute and cost for routine scans.

Cisco contrasted Antares with two common approaches: large closed commercial models, which can be costly and may conflict with data residency rules, and general open-weight models run locally, which preserve data control but often yield higher false-positive rates. Cisco described Antares as intended to offer lower operating cost and local control while aiming to reduce false positives compared with general open-weight models.

Cisco identified potential users as universities, research institutions, nonprofits and public-sector teams that maintain important software but operate with constrained security budgets and strict data-control requirements.

In a statement, Amin Saberi, professor of management science and engineering at Stanford University, wrote: “Security can’t be a luxury good, yet advanced AI-based detection has largely belonged to organizations with frontier-scale budgets. Antares’ results change that equation: near-frontier accuracy on secure code reasoning at a fraction of the cost, fast enough to run on every commit.”

Cisco framed the release as part of a broader effort to publish open specifications, reusable security knowledge, compact deployable models and measurable benchmarks for AI-assisted security workflow.