Particle.news

Cisco Releases Antares, Small Open-Weight Models for Finding Vulnerable Source Files

The compact models are built to run on local hardware so budget-constrained teams can use AI to pinpoint likely vulnerable files without sending code to the cloud.

Overview

  • Cisco published two open-weight Antares models, Antares-350M and Antares-1B, on Hugging Face and said a larger Antares-3B will follow.
  • The models are trained to act like a human investigator by searching a repository, reading candidate files, backtracking when needed, and returning a ranked list of source files with the terminal trace that produced each result.
  • Cisco created a 500-entry Vulnerability Localization Benchmark that frames tasks as CWE-style advisories and reports Antares near the top on F1 score, with Antares-3B listed among the leaders in the company’s results.
  • On Cisco’s benchmark comparisons, the Antares family finished runs much faster and at far lower estimated cost than larger models, roughly 172 times cheaper than GPT-5.5 and about 15.2 times cheaper than the best open-weight comparison.
  • Cisco positions Antares for universities, public-sector and research teams, and small security groups to enable on-premises, always-on scanning; the release complements Cisco’s Foundry Security Spec and CodeGuard tools.