Cisco has introduced two open-source artificial intelligence models designed specifically for cybersecurity applications. The models, named Antares-350M and Antares-1B, are built to identify security vulnerabilities in software code, offering organizations a cost-effective alternative to large-scale commercial AI systems.

The models are available on Hugging Face. Cisco's central pitch focuses on efficiency and value. Developer Aman Priyanshu stated on X that the smallest model, Antares-350M, detects approximately 150 times more vulnerabilities per dollar spent compared to large AI agents such as Cognition's Devin Security Swarm.

According to Axios, Cisco's internal benchmark testing showed that Antares scanned 500 code repositories in approximately 15 minutes for under one dollar. By comparison, OpenAI's GPT-5.5 took five hours and incurred over $100 in costs for the same task.

A notable advantage of the Antares models is their ability to run locally on an organization's own infrastructure, meaning sensitive source code never leaves the company's environment. This addresses a significant concern for enterprises handling proprietary or regulated codebases, where external AI services can raise data governance, compliance, and confidentiality questions.

The release also fits a broader shift in AI security tooling toward smaller, task-specific models that can be integrated into developer workflows, code review, and vulnerability triage without requiring the infrastructure footprint of general-purpose frontier models. For security teams, the practical test will be how well these models perform on their own codebases, including false positives, missed vulnerabilities, and compatibility with existing application security processes.

According to Cisco's technical report, the models were trained on approximately 72 percent security-concept data and 15 percent code search histories, reflecting a training methodology heavily focused on cybersecurity domain knowledge.

Cisco is retaining a larger three-billion-parameter version of Antares for internal use within its own products. That version reportedly performs close to GPT-5.5 and outperforms open-source models up to 200 times its size, suggesting that task-specific training can yield strong results without the computational overhead of general-purpose large language models.

The company has also indicated it is exploring the formation of an industry consortium dedicated to open AI security tools, which could signal broader collaboration within the cybersecurity community on shared, AI-driven defense capabilities.