Editorial illustration for Cisco’s Small AI Models Outperform Larger Rivals on Cost for Vulnerability Detection
Cisco's Tiny AI Models Beat Larger Rivals on Cost
Cisco released two small open-weight AI models built to find security vulnerabilities in software code, and it's making its case on price rather than raw power. Antares-350M and Antares-1B are the names, with the numbers referring to their parameter counts, tiny by the standards of frontier language models. Cisco's argument is that for this specific job, smaller and cheaper beats bigger and more expensive.
Developer Aman Priyanshu has been testing that claim publicly on X, comparing Antares against large AI agents like Cognition's Devin Security Swarm. Cisco has run its own benchmarks too, pitting the models against GPT-5.5 on the task of scanning code repositories for flaws. The results, reported by Axios, point to a wide gap in speed and cost between Cisco's compact models and the general-purpose alternative.
There's also a privacy angle: both Antares models run locally, meaning code never has to leave a company's own servers to get scanned. Cisco has published a technical report detailing what the models were trained on, and the company is holding back a larger, three-billion-parameter version for internal use in its own security products.
Antares-350M and Antares-1B spot vulnerabilities in software code. Cisco's pitch is bang for the buck. Developer Aman Priyanshu claims on X that the smallest model catches about 150 times more vulnerabilities per dollar than large AI agents like Cognition's Devin Security Swarm.
Why this matters
Cost-per-finding is becoming the real benchmark in applied AI, and Cisco just gave the industry a number to argue over: 150 times more vulnerabilities per dollar than a large agent swarm. If that holds up outside Cisco's own tests, it changes the calculus for any team deciding whether to point a frontier model at a code repo or a specialized 350-million-parameter one. Scanning 500 repos in 15 minutes for under a dollar is the kind of figure that gets budgeted into CI pipelines, not just cited in a demo.
Founders building security tooling should treat this as a nudge toward task-specific models over general-purpose ones for narrow, high-volume jobs like vuln scanning. Researchers should ask harder questions about how Priyanshu's claim was measured and whether "vulnerabilities caught" means the same thing across Antares and Devin Security Swarm. Cisco is an interested party here, not a neutral benchmark lab, and open models still need independent testing before anyone rewrites their security stack around them.
The number is compelling. The verification is not done yet.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives AMD Opens ROCm KV Cache Tech to Upstream Projects
From Archives AMD Releases Preview of Gemma 3B LLM Container for Radeon GPUs
Common Questions Answered
What are Antares-350M and Antares-1B models designed to do?
Antares-350M and Antares-1B are small open-weight AI models developed by Cisco specifically built to find security vulnerabilities in software code. These models are designed to demonstrate that smaller, more cost-effective AI solutions can outperform larger language models for specialized security tasks like vulnerability detection.
How much more cost-effective is Antares-350M compared to large AI agents like Devin Security Swarm?
According to developer Aman Priyanshu's public testing on X, Antares-350M catches approximately 150 times more vulnerabilities per dollar than large AI agents like Cognition's Devin Security Swarm. This dramatic cost efficiency difference is central to Cisco's argument that smaller models provide better value for vulnerability detection tasks.
Why does Cisco argue that smaller models are better than larger ones for vulnerability detection?
Cisco's argument is based on cost-per-finding efficiency rather than raw computational power. For the specific job of vulnerability detection, smaller models like Antares deliver superior performance metrics at a fraction of the cost, making them more practical for real-world deployment in CI pipelines and code scanning workflows.
What practical scanning capability does Cisco highlight for its small models?
Cisco highlights that its small models can scan 500 repositories in 15 minutes for under a dollar, demonstrating the cost-effectiveness that makes these models practical for integration into continuous integration pipelines. This metric represents the kind of efficiency that changes the calculus for teams deciding between frontier models and specialized smaller models.
Further Reading
- Introducing Antares: Highly Efficient Open Weight AI Models for Vulnerability Localization - Cisco Blogs
- Cisco bets on small AI for cybersecurity - Axios
- Cisco open-sources Antares AI models for vulnerability detection - DeveloperTech
- Cisco's open-weight Antares models make vulnerability localization cheaper - Help Net Security
- Cisco just launched two cyber-focused small language models, Antares-350M and Antares-1B, aimed to supercharge codebase analysis - ITPro
