Cisco has released Antares, a family of open-weight AI models built to pinpoint where known vulnerabilities live inside a codebase. Two models, Antares-350M and Antares-1B, are now available on Hugging Face to verified cyber defenders, with a larger Antares-3B model in the pipeline that Cisco plans to fold into its own security products rather than release openly.
Unlike general-purpose coding models, Antares is purpose-built for vulnerability localization: given a description of a known flaw, it searches a repository, inspects candidate files, and narrows in on the ones most likely to contain the issue, in a process meant to mirror how a human investigator would work through the same task.
The efficiency gains are the headline. Cisco says a full evaluation across 500 repositories can run in about 15 minutes on a single GPU for under a dollar, versus roughly five hours and more than $100 using GPT-5.5. That puts Antares at roughly 172 times cheaper than GPT-5.5 and about 15 times cheaper than the best-performing open-weight model Cisco tested, GLM-5.2.
On Cisco’s new Vulnerability Localization Benchmark, a 500-entry dataset built specifically for this task, the yet-to-be-released Antares-3B scored close to GPT-5.5 despite its far smaller size, while Antares-1B posted the highest recall of any model evaluated.
Because the models are small enough to run locally or on-premises, Cisco is pitching Antares at universities, public-sector institutions, and smaller security teams that can’t justify the cost of token-heavy commercial models, and at organizations that need to keep sensitive source code off the cloud entirely, including air-gapped environments.
Antares comes out of Cisco Foundation AI and joins two earlier releases aimed at the same problem space: Foundry Security Spec, a blueprint for agentic security evaluation, and CodeGuard, a secure-coding rules corpus.
Leave a Reply