antares-350m

Antares-350M Antares-350M is an open-weight, 350-million parameter language model specialized for vulnerability localization in real-world codebases. Built on IBM Granite 4.0 350M, it is trained to autonomously navigate source code repositories through a terminal interface — issuing shell commands, reading files, and iteratively narrowing in on vulnerable files — using a two-stage pipeline: supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO). Unlike traditional static analysis tools or retrieval-based approaches, Antares-350M operates as a terminal agent: it explores repositories by executing commands (grep, find, cat, and standard Unix utilities), reasons about the output, and submits a list of files it believes to contain the reported vulnerability. Antares-350M is part of the Antares family of models (350M, 1B, 3B), all built on IBM Granite 4.0 and trained with the same two-stage pipeline. The 350M model achieves a File F1 of 0.135 on the Vulnerability Localization benchmark (VLoc Bench), outperforming models many times its size including Gemini 2.5 Flash, Gemma-4-31B, GPT-5 Mini, and Qwen3.5-122B-A10B. Within the family, Antares-350M can serve as a speculative draft model for the larger Antares variants in compatible deployments: it proposes candidate tokens that a 1B or 3B model verifies, accelerating inference while preserving the larger model's localization quality. It can also run standalone in resource-constrained deployments. Antares enables organizations to deploy compact, on-premises vulnerability localization capabilities without relying on cloud-based AI services. In our evaluation setup, Antares-350M completes the full 500-task VLoc Bench sweep in approximately 11 minutes on a single H100 GPU using 16 parallel workers. Model Name: Antares-350M Model Developer: Copyright © 2026 by Cisco Systems, Inc. All rights reserved. Cisco product group: Foundation AI Model Card Contact: https://fdtn.ai/contact Technical Report: Vijay