By CiscoLearning Objectives
- Understand what vulnerability localization is and why small, specialized models fit the task
- See how open-weight, on-premises security models change the cost equation versus frontier APIs
- Evaluate when Antares is the right tool and what its limits are
What Is Cisco Antares?
Cisco Antares is a family of open-weight small language models from Cisco Foundation AI — Cisco's security-focused AI research and engineering group — released on July 21, 2026. The models do one job well: vulnerability localization. Given a description of a known security flaw and a code repository, Antares pinpoints exactly which files contain the vulnerability, so security teams can go straight to the code that needs fixing.
Two versions are available now on Hugging Face — Antares-350M and Antares-1B (350-million and 1-billion parameters) — both under the permissive Apache-2.0 license. Cisco is holding back a more capable Antares-3B (3-billion-parameter) model, which it plans to integrate into its own security products rather than release openly. Access to the open models is offered to users Cisco verifies as cyber defenders.
💡Key Concept
Small models, load-bearing AI: Antares is a deliberate bet that a tightly-scoped, security-tuned model beats a giant general-purpose model on a narrow task. On Cisco's own Vulnerability Localization Benchmark, the tiny models matched or beat about a dozen much larger open and closed systems on accuracy — while running at a tiny fraction of the compute cost. The learned model is the product's core value: cheaper, faster, and self-hostable vulnerability triage.
✅Tip
Get Antares: the open-weight models are on Hugging Face under Cisco's Foundation AI org (huggingface.co/fdtn-ai); read the technical writeup on the Cisco blog.
Pricing
- Open-weight download on Hugging Face
- Self-host on-premises
- Verified cyber-defender access
- More capable variant
- Integrated into Cisco security products
- Availability via Cisco
The open-weight models are free under Apache-2.0, and because they run on-premises there is no per-query API bill — the cost is your own compute, which is the entire point. The larger Antares-3B is not an open release; Cisco is folding it into its commercial security products.
Core Capabilities
Vulnerability Localization
Antares is purpose-built to answer a specific security question: given this known vulnerability, where in the codebase does it live? It analyzes vulnerability descriptions and scans code patterns to surface the files most likely to contain the flaw — turning a slow manual hunt across a large repository into a fast, automated first pass that points defenders at the right code.
Frontier Accuracy at a Fraction of the Cost
The headline claim is efficiency. On Cisco's Vulnerability Localization Benchmark — a 500-entry test measuring how accurately a model identifies vulnerable files in real-world repositories — Antares performed comparably to much larger systems including OpenAI's GPT-5.5 and Z.ai's GLM-5.2, while costing and running dramatically less:
| Model | Benchmark cost (500 repos) | Benchmark time |
|---|---|---|
| Antares (350M / 1B) | Under $1 | About 1 hour |
| Z.ai GLM-5.2 | About $12.50 | — |
| OpenAI GPT-5.5 | About $141 | About 4.5 hours |
Cisco says Antares beat roughly a dozen larger open and closed-weight models on accuracy in the same test — evidence that for a well-scoped security task, a specialized small model can outrun a general-purpose giant on cost, speed, and results.
On-Premises and Self-Hostable
Because the models are open-weight and small, organizations can run them on their own infrastructure without sending code to a cloud API. For security work — where the code being scanned is often the most sensitive asset a company owns — keeping analysis on-prem is a meaningful advantage, and it removes the recurring per-query cost of frontier APIs entirely.
📝Note
In Cisco's words: "The goal is to build toward a system where all security practitioners, regardless of on-prem or resource constraints, can effectively incorporate AI in everyday security operations," said Amin Karbasi, VP and Chief Scientist at Cisco Foundation AI.
Strengths
- Extremely cheap to run: scans 500 repositories for under one dollar in Cisco's benchmark — versus roughly $141 for GPT-5.5
- Fast: completes the benchmark in about an hour, roughly four-and-a-half times faster than GPT-5.5
- Frontier-comparable accuracy: matches or beats about a dozen larger models on vulnerability localization despite its tiny size
- Open-weight (Apache-2.0): downloadable, inspectable, and self-hostable
- On-premises: keeps sensitive code out of third-party cloud APIs
- Purpose-built: tuned specifically for security vulnerability localization, not a general chatbot repurposed
Limitations & Considerations
- Narrow scope: Antares localizes known vulnerabilities — it is not a general coding assistant, a full static-analysis suite, or a tool that invents novel exploits
- Gated access: the open models are offered to users Cisco verifies as cyber defenders, so access is not fully unrestricted
- Most capable version withheld: Antares-3B is not openly released — it ships inside Cisco's commercial security products
- Benchmark is Cisco's own: the headline results come from Cisco's Vulnerability Localization Benchmark; independent third-party evaluation is still limited this early
- Brand-new: released July 2026, so real-world track record across diverse codebases is still accumulating
Best Use Cases
| Task | Why Antares |
|---|---|
| Triaging known vulnerabilities in large repos | Purpose-built to point defenders at the exact vulnerable files fast |
| Cost-sensitive, high-volume code scanning | Under one dollar per 500-repo scan — a fraction of frontier-API cost |
| On-premises / air-gapped security work | Open-weight and small enough to self-host; sensitive code never leaves your infrastructure |
| Resource-constrained security teams | Small models run on modest hardware, no frontier-API budget required |
When to choose alternatives:
- General-purpose secure coding help → a frontier model like GPT-5.6, Claude, or Gemini 3.5 Flash Cyber
- Full application security platform → Cisco's own AI Defense / Hypershield or a dedicated SAST/DAST suite
- Discovering unknown/novel vulnerabilities → specialized fuzzing and exploit-research tooling
Getting Started
- Read the technical writeup on the Cisco blog to understand the model's scope and benchmark
- Find the open-weight models under Cisco's Foundation AI org on Hugging Face (huggingface.co/fdtn-ai) — start with Antares-1B for accuracy or Antares-350M for the lightest footprint
- Complete Cisco's cyber-defender verification to obtain access
- Deploy on your own infrastructure and point it at a repository with a known vulnerability description
- Compare its file-level findings against your existing triage process to calibrate trust before scaling up
- Fold it into your security operations as a fast, cheap first-pass localizer ahead of deeper manual review
Key Takeaways
- Cisco Antares is a family of open-weight small language models from Cisco Foundation AI, released July 21, 2026, built specifically to localize known security vulnerabilities inside real codebases
- The 350-million and 1-billion-parameter models are Apache-2.0 licensed and run on-premises; a more capable Antares-3B is withheld for Cisco's commercial products
- In Cisco's benchmark, Antares scanned 500 repositories for under one dollar in about an hour, versus roughly $141 and 4.5 hours for GPT-5.5 — while matching or beating about a dozen larger models on accuracy
- The bet is that a specialized small model beats a general-purpose giant on a narrow security task across cost, speed, and results — and does it without shipping sensitive code to a cloud API
- Access to the open models is offered to verified cyber defenders, and the tool is narrow by design: it localizes known flaws, it is not a general coding assistant
