Cisco releases Antares open-weight models for vulnerability localization
Cisco released Antares-350M and Antares-1B on July 21 as open-weight models designed to locate files associated with known vulnerability classes inside software repositories. Cisco says the models can run locally and that specialization made them competitive with much larger systems on its 500-task benchmark; independent Axios reporting says downloads are limited to verified defenders and the stronger 3B model remains unreleased.
Cisco on July 21 introduced Antares, a family of compact language models trained to locate vulnerable files in software repositories. The company released Antares-350M and Antares-1B through Hugging Face, while keeping a more capable 3-billion-parameter model out of public release.
What Cisco shipped
Antares is aimed at vulnerability localization rather than general code generation or automatic patching. Given a Common Weakness Enumeration category and read-only terminal access to a repository, the models search files, inspect likely paths, revise their approach, and return ranked file locations with an exploration trace.
Cisco says the smaller models can run inside an organization's own environment, reducing the need to send proprietary source code to an external model provider. The company also positions Antares as an aid to early triage, not a replacement for dependency analysis, secret scanning, dynamic testing, remediation workflows, or security review.
Axios reports that Cisco will restrict downloads to users it verifies as cyber defenders. The report also says Cisco plans to use the unreleased Antares-3B model in its own products.
What the benchmark shows
Cisco evaluated the family on its Vulnerability Localization Benchmark, which contains 500 tasks drawn from 290 repositories and 147 CWE categories. Models receive a repository with a known vulnerability and must identify the affected files; a second phase tests whether they avoid false alarms after the vulnerability has been patched.
The company's technical report says Antares-3B reached a File F1 score of 0.223, near the 0.229 reported for GPT-5.5 under the same setup. It also says 190 of the 500 tasks defeated every evaluated system. Those figures come from a Cisco-created benchmark and report, not an independent replication, and the closest-performing Antares model is not part of the public release.
Why it matters for security teams
The release is a practical example of task-specific post-training competing with model scale for a narrow security workflow. Local inference may be useful where code cannot leave a controlled environment, and a small model could make repeated repository scans less expensive.
For practitioners, the benchmark's unsolved-task rate is the more important guardrail. Antares can prioritize where analysts look, but teams still need reproducible testing, human validation, and established application-security controls before treating any finding or clean result as reliable.
Key Points
- 1Cisco released Antares-350M and Antares-1B as open-weight models for repository-level vulnerability localization, while withholding the stronger 3B version.
- 2Cisco's 500-task benchmark spans 290 repositories and 147 CWE categories, but 190 tasks remained unsolved by every evaluated system.
- 3The models can run locally and return ranked vulnerable-file candidates, but Cisco presents them as triage aids rather than replacements for established security testing and review.
Scoring Rationale
Cisco released two usable compact security models with a public technical report and benchmark, giving application-security teams a concrete local-inference option. Impact is tempered because benchmark results are company-produced, the strongest 3B model is withheld, and many tasks remain unsolved.
Sources
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
