Cisco Releases Antares, Small Language Models Specialized for Vulnerability Localization
Cisco has released Antares, a series of open-source small language models (currently available in 350M and 1B parameter versions) designed specifically for vulnerability localization. In a benchmark test, the Antares-1B model achieved a 22.4% recall rate in detecting vulnerability files, slightly outperforming GPT-5.5. This release demonstrates that highly specialized, lightweight models can match or exceed the performance of massive frontier models in specific cybersecurity tasks. Because of their small size, these models can be deployed locally on enterprise networks, offering a privacy-focused and cost-effective solution for secure code analysis. Antares is not designed to discover unknown vulnerabilities autonomously; instead, it assists security teams by using Common Weakness Enumeration (CWE) descriptions to search codebases and generate reports. Cisco notes that the model is intended as an auxiliary tool and must be integrated with broader security practices like software bill of materials (SBOM) and manual reviews.
## BACKGROUND
Vulnerability localization refers to the process of identifying the exact files or lines of code where a software vulnerability exists. Common Weakness Enumeration (CWE) is a community-developed list of common software and hardware security weaknesses, which serves as a standard language for identifying and discussing vulnerabilities.