Overview
Antares is a family of security small-language-models from cisco Foundation AI for agentic vulnerability-localization in real repositories. Built on IBM Granite 4.0; trained with SFT + GRPO for terminal-style repo navigation.
Recent Developments
- 2026-07-21: Antares-350M and Antares-1B open-weight on Hugging Face (
fdtn-ai/...), gated to vetted defenders; Antares-3B not publicly released (2026-07-23-cisco-antares-axios) - Accompanied by vloc-bench (500 tasks)
Warning
Performance claims are on Cisco’s VLoc Bench. Dual-use downloads are gated. Localization of known vulns ≠ general exploit discovery.
Related
- cisco-antares-open-weight-vuln-localization
- cisco
- hugging-face
- vulnerability-localization
- vloc-bench
- open-weight-models
- small-language-models