Antares 1B

Antares 1B

Quick answer: Antares 1B is an open-weight 1B-active-parameter security small language model by Cisco Foundation AI, purpose-built for terminal-based vulnerability localization. It achieves a File F1 of 0.209 on VLoc Bench — outperforming GLM-5.2 (753B), Gemini 3 Pro, and GPT-5 Mini — with a 128K context window and Apache 2.0 license. It completes a full 500-task repository sweep in ~13 minutes on a single H100.

At a Glance

Where Antares 1B leads

• 0.209 File F1 on VLoc Bench, beating frontier models including GLM-5.2 (753B), Gemini 3 Pro, GPT-5 Mini, and Qwen3.5-122B
• 128K context window — the largest in the Antares family — handles medium-to-large repositories
• Runs on a single GPU; full 500-task sweep in ~13 minutes on H100 with 16 workers
• Apache 2.0: permissive open-weight license, freely deployable on-premises

Where it lags

• Still below Antares 3B (0.223 F1) and frontier-cost models like GPT-5.5 (0.221 F1)
• Specialized for vulnerability localization only; not a general instruction-following model
• Gated on HuggingFace — requires agreeing to share contact info with Cisco
• Knowledge cutoff April 2025 — unaware of vulnerabilities and CWE patterns introduced after

Best for: Security teams that need a deployable, on-premises vulnerability triage agent that can handle larger codebases (up to 128K context) while running locally to keep sensitive source code off the cloud.

What Antares 1B Is

Antares 1B is the mid-tier model in Cisco Foundation AI's Antares family — a trio of open-weight security small language models (350M, 1B, 3B) designed exclusively for one problem: given a CWE identifier and a source repository, autonomously locate the files most likely to contain the vulnerability. Rather than relying on external vector databases, it operates as a terminal agent: iteratively issuing shell commands (grep, find, cat), reasoning over outputs, revising its search strategy when a path is unproductive, and ultimately submitting a ranked list of suspect files for analyst review.

Built on IBM Granite 4.0 1B via a two-stage training pipeline — SFT on cybersecurity reasoning and terminal-navigation trajectories, followed by GRPO reinforcement learning over complete multi-turn agent rollouts — Antares 1B learned to search unfamiliar codebases the way an experienced security analyst would: follow code paths, cross-reference naming conventions, inspect call graphs, and backtrack efficiently.

At ~2B total parameters with a 128K context window, Antares 1B sits between the ultra-compact 350M variant and the higher-accuracy 3B. For most practical security workflows — CI/CD triage, advisory-driven CWE investigations, and shift-left scanning — it offers the best balance in the family: frontier-grade localization accuracy at a fraction of frontier inference cost, deployable on a single GPU without sending source code to any cloud API.

Specifications

Pricing

Open-weight model under Apache 2.0. Access requires agreeing to share contact information with Cisco Foundation AI via the HuggingFace model page gating form. No per-token fees; running costs depend on your own infrastructure.

Related Models & Benchmarks

• Family: Antares 350M · Antares 3B (coming soon)
• Base model: IBM Granite 4.0 1B
• Benchmarks: VLoc Bench · SWE-bench Verified

Antares 1B vs Alternatives

Cisco's compact VLoc-Bench models. Antares 1B (20.9%) outperforms 350M (13.5%) on spatial reasoning.

Frequently Asked Questions

Scores from Benchgen evaluations. Last updated 2026-07-24.