Cisco Foundation AI Releases Antares: 350M and 1B Open-Weight Models That Localize Known Vulnerabilities Inside Real Codebases

Cisco Foundation AI has unveiled Antares, a groundbreaking family of open-weight language models specifically engineered to identify and localize known security vulnerabilities within real codebases. This development represents a significant advancement in AI-driven cybersecurity, offering developers and security teams powerful tools to detect and remediate code vulnerabilities more efficiently than previous solutions.

Cisco Foundation AI released two open-weight models in the Antares family: a 350 million parameter model and a 1 billion parameter model. The larger Antares-1B model achieved a File F1 score of 0.209 on the newly introduced Vulnerability Localization Benchmark, demonstrating superior performance compared to significantly larger models. Remarkably, Antares-1B outperformed GLM-5.2, which operates at 753 billion parameters, and Google's Gemini 3 Pro. This efficiency breakthrough demonstrates that specialized, smaller models can exceed the capabilities of general-purpose large language models when optimized for specific security tasks.

The models were trained specifically to locate where known vulnerabilities exist within actual codebases, addressing a critical pain point in the software development lifecycle where identifying vulnerable code sections remains time-consuming and resource-intensive.

• Cost-effective security scanning: Smaller models reduce computational overhead, enabling organizations to deploy vulnerability detection on-premises or in resource-constrained environments
• Faster vulnerability patching: Precise localization accelerates the remediation process by eliminating manual code review requirements
• Open-weight accessibility: Availability as open models democratizes advanced security capabilities for smaller teams and organizations
• Benchmark standardization: The Vulnerability Localization Benchmark establishes new metrics for evaluating code security models
• Competitive pressure: Performance advantages over larger models may reshape AI model selection criteria across the industry

The release of Antares addresses a critical gap in cybersecurity tooling where manual vulnerability identification remains a bottleneck. By combining specialized training with efficient model architecture, Cisco Foundation AI has demonstrated that targeted AI applications can achieve superior results compared to general-purpose alternatives. This development will likely influence how organizations approach code security and accelerate adoption of AI-powered vulnerability management systems across the industry.