Antares 350M

Antares 350M

Quick answer: Antares 350M is an open-weight 350M-active-parameter security small language model by Cisco Foundation AI, purpose-built for terminal-based vulnerability localization. Despite its compact size, it achieves a File F1 of 0.135 on VLoc Bench — outperforming Gemini 2.5 Flash, Gemma 4 31B, GPT-5 Mini, and Qwen3.5-122B — while running the full 500-task sweep in ~11 minutes on a single H100. Apache 2.0 licensed and free to download.

At a Glance

Where Antares 350M leads

• Outperforms models 80–350× larger on VLoc Bench (Gemini 2.5 Flash, Gemma 4 31B, Qwen3.5-122B-A10B)
• Tiny footprint enables local/on-premises deployment with no cloud exposure of sensitive code
• Can serve as a speculative draft model for Antares 1B/3B in compatible inference stacks
• Apache 2.0: no usage restrictions beyond Cisco's contact-info access form on HuggingFace

Where it lags

• 32K context window limits performance on large repositories (>10MB codebases)
• Weakest in the Antares family — 0.135 File F1 vs. 0.209 for Antares 1B
• Specialized for vulnerability localization only; not a general instruction-following model
• Knowledge cutoff April 2025 — blind to newer CVE/CWE patterns

Best for: CI/CD pipelines and resource-constrained security environments needing a lightweight, local vulnerability triage agent that keeps source code on-premises.

What Antares 350M Is

Antares 350M is the smallest model in Cisco Foundation AI's Antares family, a trio of open-weight security small language models (350M, 1B, 3B) trained for one narrow task: given a CWE identifier and a source repository, find the files most likely to contain the reported vulnerability. It operates as a terminal agent — issuing shell commands (grep, find, cat), reasoning about the output, and iteratively narrowing to a ranked file list — without any external vector database or retrieval system.

Built on IBM Granite 4.0 350M via a two-stage pipeline (SFT on cybersecurity reasoning and terminal-navigation data, then GRPO reinforcement learning over multi-turn agent trajectories), Antares 350M learned to mimic how a human investigator searches an unfamiliar codebase: follow naming conventions, inspect call paths, compare candidate files, and backtrack when a path is unproductive.

At 0.4B total parameters and a 32K context window, Antares 350M is the fastest and lowest-VRAM option in the family — completing the 500-task VLoc Bench sweep in approximately 11 minutes on one H100 with 16 parallel workers — making it practical for automated CI/CD triage gates and edge/on-premises security deployments.

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

Antares 350M 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.