Cisco Foundation AI Releases Antares: 350M and 1B Open-Weight Models That Localize Known Vulnerabilities In…

What it does

Cisco Foundation AI launched Antares, a family of open-weight small language models designed specifically to locate known security vulnerabilities within real-world codebases. Antares comes in two sizes, 350 million and 1 billion parameters, and targets the practical problem of vulnerability localization rather than general code generation or analysis. The larger Antares-1B model reaches a File F1 score of 0.209 on the new Vulnerability Localization Benchmark, outperforming models like GLM-5.2, which has 753 billion parameters, as well as Gemini 3 Pro. This shows a focused training approach can deliver better precision in niche tasks at dramatically smaller model sizes.

Why it matters

Pinpointing where vulnerabilities live inside code is a critical step for security teams and developers aiming to fix issues quickly. Antares’s ability to localize known weaknesses efficiently can cut down on the time and computing resources needed for scanning large codebases. With a 500-task evaluation running in about 13 minutes on a single Nvidia H100 GPU for under a dollar, this represents a cost-effective way to integrate vulnerability detection into development cycles or security audits. The model’s strong performance over much larger but more general language models also pressures vendors to deliver specialized solutions rather than purely scale.

Who it is for

Antares targets developers, security engineers, and toolmakers who need to identify vulnerabilities within their existing software projects quickly and accurately. It can be incorporated into automated code review tools, CI/CD pipelines, or vulnerability management platforms. For businesses facing rising costs from long vulnerability turnaround times or external audits, Antares offers a more resource-efficient alternative. Smaller teams and startups with limited cloud budgets may also benefit from a model that delivers competitive accuracy without massive infrastructure.

The catch

As a newly released open-weight model, Antares’s real-world effectiveness depends on integration into developer workflows and toolchains. Since the core benchmark focuses on known vulnerabilities, the model may be less adept at spotting novel or subtle zero-day flaws. Adoption will also hinge on how easily teams can run the model at scale or tailor it to specific languages and frameworks. The Granite 4.0 checkpoints score near zero without training, underscoring the importance of fine-tuning for this use case versus off-the-shelf large models.

What to watch next

Further evaluation of Antares in live environments will show whether its benchmark promises translate into faster vulnerability remediation and improved security postures. Watch for integrations with established developer security platforms and open-source tools that can expand adoption. Follow-up releases may push parameter counts or domain-specific training to expand coverage and catch emerging threats. How competitors respond with their own compact, specialized detection models will shape the vulnerability scanning market’s next phase.

AI Quick Briefs Editorial Desk