Cisco Foundation AI Releases Antares Security Models

Cisco Foundation AI released Antares, a family of security small language models (SLMs) designed for vulnerability localization on March 14, 2024. The models are trained to identify files containing specific vulnerabilities within code repositories. Two open-weight models, Antares-350M and Antares-1B, are now available on Hugging Face under the Apache 2.0 license. Alongside the models, Cisco also released the Vulnerability Localization Benchmark (VLoc Bench), a 500-task agentic evaluation tool, also under the Apache 2.0 license.

The Antares models aim to streamline the initial triage process in software security by connecting external vulnerability knowledge, sourced from public databases and Common Weakness Enumerations, with internal source code. This task is often time-consuming for developers who must navigate large, complex codebases. While Antares does not replace existing application security tools such as dependency scanners or secret scanners, it addresses the cost concentration in the initial search phase. The Antares family includes three decoder-only transformer models with 350 million, 1 billion, and 3 billion parameters, all initialized from IBM Granite 4.0 checkpoints. The Antares-350M and Antares-1B models are publicly released, while the Antares-3B model is not yet published.

Performance benchmarks indicate that Antares-1B achieves a File F1 score of 0.209 on the VLoc Bench. This performance is competitive, with GPT-5.5 reaching 0.229 and a 753 billion parameter open-weight model scoring 0.186. The models utilize a shared tokenizer and architecture, including grouped-query attention, SwiGLU MLPs, RMSNorm, RoPE, and shared input/output embeddings. The Antares-1B model boasts a context window of 128K tokens, with 40 layers and 2048 hidden units. The development by Cisco Foundation AI focuses on specialized tasks within the AI security domain, providing accessible tools for the developer community.