Cisco Foundation AI Releases Antares: Compact Models for Code Vulnerability Localization
Cisco Systems
IBM
Cisco Foundation AI has released Antares, a family of security small language models (SLMs) for vulnerability localization. The 1B parameter model achieves 0.209 File F1 on the new VLoc Bench, outperforming some much larger models. Antares-350M and Antares-1B are open-weight under Apache 2.0 on Hugging Face.
Cisco Foundation AI released Antares, a family of security small language models (SLMs) designed specifically for vulnerability localization: given a vulnerability description and a repository, they find the files containing the flaw. Two models, Antares-350M and Antares-1B, are available open-weight under Apache 2.0 on Hugging Face. Cisco also released the Vulnerability Localization Benchmark (VLoc Bench), a 500-task agentic evaluation, under the same license. Antares models are decoder-only transformers based on IBM Granite 4.0 checkpoints. The 1B model reaches 0.209 File F1 on VLoc Bench, above GLM-5.2 at 753B parameters (0.186) and Gemini 3 Pro, while GPT-5.5 scores 0.229. The 3B model (not released) scores 0.223. Supervised fine-tuning (SFT) and reinforcement learning via GRPO provided significant gains; GRPO added 11-25% File F1 and reduced variance by 42-65%. A full 500-task sweep costs under $1 on one H100, compared to $12.50 for GLM-5.2 and $141 for GPT-5.5.
- Сокращения
- SLM = Small Language Model — малая языковая модель
- CWE = Common Weakness Enumeration — общее перечисление слабостей
- CVE = Common Vulnerabilities and Exposures — общие уязвимости и риски
- F1 = F1 Score — метрика F1
- GRPO = Group Relative Policy Optimization — групповая оптимизация политики
- SFT = Supervised Fine-Tuning — контролируемая донастройка
Source: MarkTechPost —
original