Antares-1B

Antares-1B is the larger of the two vulnerability-localization models Cisco Foundation AI released on 21 July 2026 under Apache 2.0. Its job is narrow and well defined: given a CWE description of a known vulnerability, find the files in a codebase where it actually lives. It is not a code-review chatbot and not an exploit generator. The interesting part is how it works. Antares is a terminal-based agentic model — it explores a mounted repository with ordinary shell commands such as grep, find and cat, follows an iterative search pattern, and returns ranked candidate files together with the trace of what it looked at, so a human reviewer can check the reasoning rather than trust a verdict. Cisco trained it in two stages: supervised fine-tuning on cybersecurity reasoning, deep-research data and code-search trajectories, then GRPO reinforcement learning over complete agent trajectories with multi-component verifiable rewards. Architecturally it is an auto-regressive decoder-only transformer built on IBM's Granite 4.0 1B backbone: 40 layers, hidden dimension 2048, 16 attention heads with 4 KV heads (grouped-query attention), SwiGLU, RMSNorm, RoPE, a 100,352-token vocabulary and a 128,000-token context window. It clears the full 500-task VLoc Bench sweep in roughly 13 minutes on a single H100 with 16 parallel workers, and Cisco's benchmark table shows it and its 350M sibling beating a dozen much larger models on this specific task at a fraction of the cost. The practical argument is deployment, not leaderboards. A model this size runs on-premises, in CI/CD, in environments where source code is not allowed to leave the building — which is exactly where cloud security assistants cannot go. Cisco is explicit that Antares is a triage tool, not a replacement for a full application-security toolchain, and that it should run sandboxed with human oversight. A 3B variant is announced but not yet released.

▸News

▸Videos

No videos yet.

▸Reviews