Antares is a suite of small, open-weight language models developed by Cisco Foundation AI to address a precise security challenge: localizing known vulnerabilities within internal codebases. With Antares you input a vulnerability description—typically a Common Weakness Enumeration (CWE)—and the model performs autonomous, multi-step reasoning over your repository (via terminal-like access) to return files most likely to contain the issue. Compact models in this family make it feasible to run security workflows locally—avoiding cloud exposure—while offering benchmarks close to the largest, frontier-scale AI systems at significantly lower cost.
Key Features
Antares includes models at three sizes—350 million, 1 billion, and 3 billion parameters—each built on a shared architecture and training pipeline. The 350M and 1B versions are currently released as open-weight; the 3B model is to follow soon.
These models follow an agentic workflow: starting with vulnerability descriptions, they iteratively explore codebases—reading files, gathering evidence, switching strategies when needed—and deliver a ranked list of source files likely to contain the vulnerability. They also return the exploration trace, aiding analyst review.
Performance metrics come from VLoc Bench, a 500-task benchmark covering 290 repositories and 147 CWE categories. Antares-3B approaches the performance of a frontier model (GPT-5.5), while Antares-1B and Antares-350M outperform many much larger open-weight models, especially in cost versus effectiveness.
Efficiency is a core selling point: Antares-3B completed the full 500-task evaluation in about 15 minutes on a single high-end GPU, at less than $0.002 per task. The smaller models offer even lower requirements—making them suitable for edge, on-premises, or resource-constrained settings.
Antares comes with a command-line interface. Users can target specific CWE checks, preview results, inspect reports in JSON, markdown, or SARIF format, and perform repository sweeps. This makes integration into existing CI/CD pipelines, advisory triage, or security workflows more straightforward.
Who is it for?
Antares is designed for teams and organizations with strong security needs but limited infrastructure or strict privacy requirements. The primary audiences include:
Security teams in universities, public-sector institutions, nonprofits, or other organizations with constrained compute resources.
Enterprises or development groups operating under compliance rules or data privacy mandates that discourage sending source code to external APIs or third-party cloud services.
DevSecOps practitioners who need fast, continuous scanning—ideally on every commit—to catch known vulnerabilities early, especially as codebases evolve or when integrating third-party advisory information.
Consultants, service providers, or partners offering managed security, vulnerability assessments, or secure code review workflows. Antares enables them to build services around localized vulnerability detection, advisory triage, and remediation planning without depending on expensive frontier models.
Pricing
There is no subscription or licensing fee disclosed. Antares-350M and Antares-1B are available now as open-weight models under an Apache-2.0–compatible license. Users can access them via model repositories. In benchmarking, the evaluation of 500 vulnerability localization tasks with Antares-3B cost approximately 82 cents and took about 15 minutes on a single high-end GPU; smaller models incur lower inference costs. No pay-per-use fees or hosted API pricing are stated officially.
Final thoughts
For decision-makers evaluating tools to improve code security workflows, Antares presents a compelling option in scenarios where source control, privacy, and cost are primary constraints. Its open-weight release of 350M and 1B models makes it accessible and transparent; organizations can inspect, self-host, and integrate without depending on external APIs. The performance metrics offer strong evidence that task-specific, compact models trained for vulnerability localization can rival much larger, generalist models for this specific task.
However, Antares does not aim to replace full security toolchains. It does not cover dependency scanning, dynamic testing, threat modeling, or remediation execution. Instead, it functions as a precision tool for initial vulnerability signal triage—helping teams focus analysts’ attention on code files most likely to contain issues. For many organizations—especially those constrained by compliance, privacy, or budget—such focused automation can make a meaningful difference.
Visit the official website for more.
Keep up to date with our stories on LinkedIn, Twitter, Facebook and Instagram.