Cisco Launches Antares Open-Weight AI Models for Vulnerability Localization

Cisco has launched Antares-350M and Antares-1B, two cyber-focused small language models built to identify vulnerable files in source-code repositories from vulnerability descriptions. The company said the open-weight models will be released on Hugging Face and are designed to run locally, allowing security teams to analyze sensitive codebases without sending source code to cloud-hosted frontier models. Cisco positioned Antares for repository-level vulnerability localization, advisory triage, static-analysis augmentation, and shift-left security workflows in CI/CD pipelines.

Cisco said Antares mimics a human investigator by tracing relevant code patterns, reviewing candidate files, and narrowing findings as new evidence emerges. In its announcement and related reporting, Cisco claimed the models can perform repository analysis at a fraction of the compute cost of larger general-purpose models, citing a 500-entry evaluation completed in 15 minutes on a single GPU for less than $1. Cisco also said Antares outperformed several OpenAI and Google models on vulnerability-finding benchmarks, though reporting noted it did not exceed OpenAI's premium GPT-5.5 (xhigh) and that Anthropic's Claude Opus 4.6 scored higher in a Cisco technical report.