Cisco Systems on July 21 introduced Antares, a family of small language models built to pinpoint exactly where known security vulnerabilities sit inside a codebase, and released the first two as open-weight downloads on Hugging Face. The launch, covered by SiliconANGLE, Axios, and Help Net Security, marks one of the most concrete attempts yet to bring specialized small models into a corner of software security that has long been slow, expensive, and reliant on human investigators.
Antares comes from Cisco Foundation AI, the company's research and engineering group focused on security-specific artificial intelligence. The project tackles a problem known as vulnerability localization, the work of connecting external vulnerability data, such as public databases, advisories, and Common Weakness Enumeration (CWE) entries, to the specific files in a repository where a flaw is most likely to live. To follow the latest developments in this space, see our AI industry coverage.
Small Models Built for a Specific Job
Two models are available now: Antares-350M and Antares-1B, with a larger Antares-3B model due to follow. All are small enough to run locally, which Cisco said lets security teams keep sensitive source code inside their own environment rather than sending it to a cloud service. That is a notable selling point for universities, public-sector bodies, nonprofits, and smaller security teams that lack the budget for token-heavy commercial frontier models.
Rather than scanning a repository with fixed rules, Antares works through it the way a human investigator would. Each model starts from a vulnerability description, searches for relevant code patterns, reads candidate files, folds in new evidence as it goes, changes direction when a path leads nowhere, and narrows toward the files most likely to matter. The output is a ranked list of source files likely to contain the relevant vulnerability, along with the terminal exploration trace that produced it.
A Custom Benchmark and Dramatic Cost Savings
To measure the models, Cisco built its own test, the Vulnerability Localization Benchmark, a 500-entry set that requires a model to navigate an unfamiliar codebase and recognize vulnerability patterns tied to specific weakness categories. Cisco argued that existing code-search benchmarks measure whether an agent can find code relevant to a general software issue, not whether it can locate vulnerable files from security descriptions. The closest comparable work the company cited is a system called CodeScout, which showed that code localization can be trained and measured with a standard Unix terminal.
On that benchmark, Cisco said the Antares models beat roughly a dozen larger open- and closed-weight models on accuracy while running far cheaper. The company put the cost of an Antares evaluation at under one dollar, against roughly $12.50 for the strongest open-weight model it tested, GLM-5.2, and about $141 for the strongest closed-source model, GPT-5.5. Those figures work out to 15.2 times and 172 times cheaper, respectively. Antares also finished benchmark runs in about an hour, compared with roughly four and a half hours for GPT-5.5.
The performance gap highlights a broader industry trend: for narrowly defined tasks, carefully trained small models can match or exceed general-purpose frontier models at a sliver of the compute cost. As Axios reported, Cisco is effectively betting that smaller, purpose-built AI will win in enterprise cybersecurity, where teams care more about reliability, cost, and data privacy than raw model size.
Part of a Broader Security AI Push
Antares is the latest piece of a wider Cisco effort around AI in security. In October, the company released Project CodeGuard, an open-source framework of secure-coding rules for AI coding agents. In May, it published the Foundry Security Spec, an open blueprint for building agentic security evaluation systems. Antares extends that line of work by putting a usable, open-weight model directly into the hands of practitioners.
The release also arrives amid growing concern about AI-related software risks. The same month Cisco published the Foundry Security Spec, Bloomberg reported that an advanced Anthropic model triggered an emergency meeting between Treasury Secretary Scott Bessent, Federal Reserve Chair Jerome Powell, and bank executives, underscoring how seriously regulators are treating the security implications of powerful AI systems.
What It Means for Developers and Security Teams
For organizations that have hesitated to adopt AI-driven code analysis because of cost or data-exposure concerns, Antares offers a middle path. Because the models are open-weight and small enough to run on modest hardware, teams can deploy them behind their own firewalls, auditing every inference call without paying per-token fees to a cloud provider. The open-weight licensing also means researchers can inspect, fine-tune, and extend the models for their own domains.
The trade-off is that small models are specialized by design. Antares will not draft a marketing email, summarize a meeting, or answer a general knowledge question. What it is built to do, locate the precise files where a known class of vulnerability is likely hiding, it appears to do better than models a thousand times its size, at least on Cisco's own benchmark.
Whether independent researchers confirm those results on their own codebases will determine how widely Antares is adopted. For now, the launch reinforces a message that has been building across the AI industry all year: when the task is narrow enough, smaller can be smarter, cheaper, and safer.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
