Cisco Foundation AI Introduces Antares: Compact Models for Vulnerability Detection in Code

Cisco Foundation AI has announced Antares, a family of small language models (SLMs) specialized in the narrow task of vulnerability localization. Two models with open weights, Antares-350M and Antares-1B, are already available on Hugging Face under the Apache 2.0 license; a third model, Antares-3B, has not yet been released. Along with them, the benchmark VLoc Bench (500 tasks) has been published under the same license. The models operate in an agentic loop: they receive a description of the vulnerability type (CWE) and can execute up to 15 commands in an isolated Docker container, exploring the repository. Results show that Antares-1B achieves a File F1 of 0.209, which is higher than that of the GLM-5.2 model with 753 billion parameters (0.186) and Gemini 3 Pro. The strongest version, Antares-3B (0.223), slightly lags behind GPT-5.5 (0.229). All three models are initialized from base checkpoints of IBM Granite 4.0, which yield zero results without fine-tuning. The main gain comes from supervised fine-tuning, and then GRPO adds 11-25% File F1 and reduces result variance by 42-65%. A full run of 500 tasks costs less than one dollar on a single H100. Cisco emphasizes that Antares does not replace existing security tools but merely automates the first stage of triage.