TL;DR — Key Takeaways
- Cisco’s Antares models specialize in locating vulnerable code rather than serving as general-purpose coding assistants.
- The first models are Antares-350M and Antares-1B, with a larger Antares-3B model planned.
- Cisco says a 500-entry evaluation cost less than $1 and took roughly 15 minutes on a single GPU.
Cisco has released Antares, a new family of cheap, local security AI models to help you find vulnerable code without the cost and data exposure that often come with large cloud-based models.
You probably don’t think of Cisco as an AI company. I know I don’t. But that just goes to show I don’t know as much as I should. The latest proof of that is its release of Antares. This is a new family of small language models designed to pinpoint vulnerable code inside large software repositories for less money than using frontier-scale AI systems.
The first two models, Antares-350M and Antares-1B, will be released as open-weight models on Hugging Face, with a larger Antares-3B model coming later.
Antares’s job is to cut down on the ever-growing costs of using high-end frontier models such as Anthropic’s Claude Mythos 5. Rather than competing head-on with broad coding assistants, Antares specializes in finding vulnerabilities. That’s it. That’s all. That may be enough for many cost-conscious developers.
Cisco claims that in its internal testing, a full 500-entry evaluation run with Antares took roughly 15 minutes on a single GPU and cost less than $1. That was about 15 times cheaper than the best open model Cisco compared it against and 172 times cheaper than the leading frontier model in the evaluation.
To measure that performance, Cisco created what it calls the Vulnerability Localization Benchmark, a 500-entry dataset focused specifically on security localization rather than general coding ability. On that benchmark, Cisco says Antares-3B achieved a File F1 score of 0.223, close to GPT-5.5 at 0.229 despite being far smaller, while Antares-1B posted the highest recall among the evaluated systems and Antares-350M outperformed several much larger general-purpose models.
In addition, you can run Antares locally or on-premises. This enables your security teams to keep sensitive source code inside your own environments. That makes the models suitable for air-gapped and closed-network deployments and could appeal to organizations that have been reluctant to send proprietary code into cloud AI services.
Cisco is framing Antares as part of a broader push to make AI-based security tooling more accessible beyond the largest enterprises. In the embargoed announcement, the company says compact models such as Antares could help universities, public-sector institutions, nonprofits, and smaller security teams adopt AI-assisted code analysis without high-end frontier model costs.
“Security can’t be a luxury good, yet advanced AI-based detection has largely belonged to organizations with frontier-scale budgets,” said Amin Saberi, professor of management science and engineering at Stanford University, in a statement. “Antares’s results change that equation: near-frontier accuracy on secure code reasoning at a fraction of the cost, fast enough to run on every commit.”
Cisco added that Antares is part of a larger security AI effort that also includes Foundry Security Spec. This is a model-agnostic framework for agentic security evaluation. Cisco is also offering CodeGuard. This is an open-source, model-agnostic security framework that embeds secure-by-default practices into AI coding agent workflows.
Put it all together, and Cisco argues that practical AI security will depend less on one-off demos and more on measurable, governable systems that defenders can run in real-world environments.
Recent Articles By Author
Frequently Asked Questions
What is Cisco Antares?
Antares is a family of compact AI models designed specifically to identify and locate security vulnerabilities in software repositories.
How is Antares different from general coding assistants?
Antares is narrowly focused on vulnerability localization. It is not intended to compete with broad coding assistants that generate code, answer programming questions or perform a wide range of development tasks.
What Antares models are available?
Cisco has announced Antares-350M and Antares-1B as open-weight models, with a larger Antares-3B model expected later.