FULL STORY

Cisco Releases Open-Source Antares for Code Vulnerability Localization

Cisco launched the Antares family of open-weight small language models, designed specifically for code vulnerability localization.

2026-07-21 ~ 2026-07-25 · 2 episodes · 12 posts

Episode 1 · Cisco Launches Open-Source Antares Models for Code Vulnerability Localization (2026-07-21, 8 posts)

Cisco has officially launched the Antares series, a set of open-weight small language models, marking a strategic move into the AI model space. Unlike general-purpose chatbots, Antares focuses on cybersecurity defense, specifically targeting the efficient localization of known vulnerabilities within codebases. The series currently includes the Antares-350M and Antares-1B models, which are available on Hugging Face.

Confirmed

According to official announcements, Antares is designed for high efficiency in critical security tasks. Cisco claims that these models outperform many existing open-source and closed-source models in vulnerability localization. Furthermore, Antares aims to address the issue of enterprise AI budget overruns by advocating for task-specific AI. The inference costs are reportedly reduced by up to 172 times compared to other solutions, significantly lowering the financial barriers for vulnerability localization in cybersecurity.

Unconfirmed

As noted by author @shashib, Cisco is currently evaluating which groups will be granted access to the system, and the future scope of availability remains to be seen.

Why it matters

The release of Antares highlights a broader industry trend toward vertical-specific and task-oriented AI models. By offering an ultra-low-cost inference solution, Cisco aims to provide enterprises with a highly cost-effective security defense tool, which could fundamentally shift budget allocation and technology choices in code security auditing.

Episode 2 · Cisco Open-Sources Antares Small Models for Code Vulnerability Localization (2026-07-23, 4 posts)

Cisco released Antares, a family of open-weight small language models ranging from 350M to 3B parameters. Designed specifically for code vulnerability localization, these models provide security teams with an efficient, local alternative to large proprietary models.