Open SourceCiscoReleased 2026-07

Antares-1B

Efficient open-weight small language model for finding known vulnerabilities in codebases

Context window

128000

Input / 1M tokens

Free

Output / 1M tokens

Free

Provider

Cisco

Data verified 2026-07-26

Antares-1B is a specialized security small language model (SLM) designed to pinpoint where known vulnerabilities exist within codebases. Built by Cisco Foundation AI, it specializes in vulnerability localization - connecting external vulnerability knowledge to internal source code - and runs locally without requiring source code to be sent to external providers. The model significantly outperforms larger closed- and open-weight models on vulnerability detection benchmarks while running at a fraction of the cost.

Capability index

Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.

Coding
7
Reasoning
6
Math
0
Multimodal
0
Long context
5
Speed
9
Cost efficiency
10

How to access it

Download open-weight model from Hugging Face. Cisco vets access on a case-by-case basis. Models run locally on-premises.

Strengths

  • 172x cheaper than GPT-5.5 on vulnerability localization benchmark
  • 15.2x cheaper than GLM-5.2 open-weight model
  • Faster execution: 15 minutes for 500-task benchmark vs 5 hours for frontier models
  • Runs locally, maintaining data sovereignty
  • Outperforms Gemini 3 Pro on vulnerability localization
  • Specialized focus reduces false positives

Best for developers who...

Security teams with limited resourcesOrganizations requiring local code analysisContinuous vulnerability scanning with cost constraintsRepository-level vulnerability triage

When to choose it (and when not to)

Reach for Antares-1B when...

  • When cost efficiency is critical for repeated vulnerability scanning
  • When source code must remain on-premises or private
  • For initial triage phase of vulnerability investigation
  • For organizations with constrained security budgets

Look elsewhere if...

  • Antares is not a replacement for full application security stack
  • Should not replace dependency analysis, SAST, or human review
  • Not designed as a general-purpose coding assistant
  • Does not generate code or answer programming questions

How to use it

  • Provide CWE (Common Weakness Enumeration) identifiers or vulnerability descriptions
  • Use in CI/CD workflows triggered by security advisories or dependency alerts
  • Leverage ranked file list output as starting point for security analyst triage

Quickstart

Python
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = 'Cisco/antares-1b'
model = AutoModelForCausalLM.from_pretrained(model_id)
tokenizer = AutoTokenizer.from_pretrained(model_id)

Access requires Hugging Face verification from Cisco. Model runs locally with standard transformer inference.

API model id: fdtn-ai/antares-1b

Benchmarks

BenchmarkScoreNotes
Vulnerability Localization Benchmark (VLoc Bench)Beats Gemini 3 Pro; matches GLM-5.2500-task dataset evaluating vulnerability localization. Antares-1B demonstrates near-frontier accuracy at fraction of cost.
Cost efficiency$0.71 per 500-task benchmark runCompared to $12.50 for GLM-5.2 and $141 for GPT-5.5
Speed15 minutes on single Nvidia H100 GPUvs 5 hours for GPT-5.5 on same 500-repository evaluation
VLoc Bench (Vulnerability Localization)0.209File F1 score; highest recall at 0.224; outperforms GLM-5.2 (753B params) at 0.186

Source: Cisco Foundation AI Official Blog

Compare Antares-1B

Compare Antares-1B with any other model

Build a comparison →
All model comparisons →

Learn the concepts