Cisco Open-Sources Antares Models for Code Vulnerability Localization

Cisco Foundation AI has introduced Antares, a family of security-focused small language models designed to identify files containing known vulnerabilities in real-world codebases. Vulnerability localization is a critical early step in software security work, helping researchers narrow their search before analyzing or patching flaws. Smaller, specialized models may also appeal to companies seeking local deployment, lower computing costs and tighter control over proprietary source code.

Cisco Foundation AI released open weights for Antares models with 350 million and 1 billion parameters, enabling developers to test and integrate them in their own environments. Benchmark results published with the release showed the 1B model approaching GPT-5.5 on vulnerability-file localization while outperforming some larger open-source models. The results suggest task-specific training can allow compact models to compete with general-purpose systems that require substantially greater computing resources.