Built and released Antares, VLoc Bench, and related tooling (Foundry Security Spec, CodeGuard); publishes the models under its 'fdtn-ai' Hugging Face org and controls the request-gate that decides who can download weights.
Benchmarked frontier closed model used as the accuracy ceiling; Antares-3B trails it only slightly (0.223 vs 0.229 File F1) while costing about 172x less per evaluation, undercutting the case for defaulting to frontier models on this task.
Maker of the 753-billion-parameter open-weight GLM-5.2, which the 1-billion-parameter Antares-1B outperforms on the benchmark while costing roughly 15x less per evaluation.
GO
Google (Gemini 3 Pro / Gemini 2.5 Flash)
Benchmarked competitor; Antares-1B is reported to outperform Gemini 3 Pro, and Antares-350M outperforms Gemini 2.5 Flash on the same benchmark.
UN
Universities, public-sector institutions, nonprofits, smaller security teams
Cisco's stated target beneficiaries who can now adopt AI-assisted vulnerability analysis without a frontier-model budget, given Antares's sub-dollar per-scan cost and local deployment.
VP at Cisco Foundation AI credited with announcing Antares via the company blog post.