Cisco today unveiled Antares, a family of open-weight small language models designed to help security teams locate software vulnerabilities more quickly while reducing computing costs and keeping sensitive source code inside enterprise environments.
The company is releasing two models, Antares-350M and Antares-1B, on Hugging Face. The models are designed specifically for repository-level vulnerability localization, a process that helps security teams identify where known vulnerabilities exist within large codebases.
Unlike general-purpose coding models, Antares is intended to run locally, allowing organizations to analyze proprietary code without sending it to cloud-based AI services. Cisco said the approach supports deployments in air-gapped and other security-sensitive environments while lowering inference costs.
Cisco said a full evaluation of 500 repositories can be completed in about 15 minutes on a single GPU for less than $1, making the workflow roughly 15 times less expensive than the best-performing open model it evaluated and 172 times cheaper than a leading frontier model.
The company also introduced the Vulnerability Localization Benchmark, a new 500-entry benchmark created to measure AI models’ ability to identify vulnerable files in software repositories. Cisco said its upcoming Antares-3B model achieved performance approaching GPT-5.5 on the benchmark despite being significantly smaller, while Antares-1B recorded the highest recall among evaluated models.
The release reflects a broader enterprise AI trend toward smaller, specialized models designed for specific business tasks rather than relying exclusively on large, general-purpose AI systems.