Cisco Antares: security models small enough to run local

Cisco's Foundation AI group released Antares, two open-weight models built for one narrow security job: given a vulnerability advisory, find which files in a repository actually contain the flaw. They're 350 million and 1 billion parameters, Apache 2.0 licensed, and small enough to run on a laptop. For anyone who has ever pasted a client's source into a chat window to answer "does this affect us," that last part is the whole story.

What actually happened

Per Cisco's announcement, Antares-350M and Antares-1B do vulnerability localization: they take a CWE description or advisory, explore a codebase through terminal commands, and return a ranked list of files likely to contain the issue. The Antares-1B model card fills in the details Cisco's blog leaves out — the models initialize from IBM Granite 4.0 checkpoints and were trained with supervised fine-tuning followed by GRPO reinforcement learning.

Cisco also published the benchmark it scored on, VLoc Bench, 500 tasks. Antares-1B posts a file-level F1 of 0.209, with precision 0.262 and recall 0.224, beating GLM-5.2 (753B parameters), Gemini 3 Pro, GPT-5 Mini, and Qwen3.5-122B on the same task. SecurityWeek reports the cost gap Cisco claims: roughly 172x cheaper than GPT-5.5 and 15.2x cheaper than GLM-5.2 for equivalent work. A 3B version exists and is being held back for Cisco's own products.

Read that F1 number honestly, because Cisco didn't hide it. 0.209 is not an oracle. It's a triage ranker that puts the right files near the top of a list a human then reads. That's a real job — it's most of the work in the first hour of a CVE response — but it is not "the AI finds your bugs."

Why small open-weight models matter for your business

The interesting property here isn't accuracy, it's shape. A 1B-parameter Apache-2.0 model is a file you download once and run forever. It doesn't have a price page, a rate limit, a deprecation notice, or a terms-of-service clause about training on your inputs. When your compliance answer to "where does our source code go" has to be "nowhere," that constraint used to mean no AI at all. Now it means a 1B model and a GPU you already own — or, at this size, just a decent CPU.

The practical use for a small shop is unglamorous and worth building: when a CVE lands in a dependency you ship, you need to know within the hour whether it touches your code paths. Today that's a person grepping. Antares turns it into a ranked short list, locally, for effectively zero marginal cost — which means you can afford to run it on every advisory instead of only the scary ones.

Two cautions. This is a localization tool, not a scanner; it needs an advisory as input, so it complements dependency scanning rather than replacing it. And the model Cisco kept — the 3B — is the one going into Cisco products. Open-weight releases from vendors are frequently the demo tier. Take the free capability, build on it, and don't confuse it with a commitment.

Sources: Cisco Blogs, Hugging Face model card, SecurityWeek.