If the most expensive part of securing a small codebase is paying a frontier API to read it for you, Cisco Foundation AI’s new Antares models remove that bill. On July 21, 2026, Cisco published Antares-350M and Antares-1B, a pair of open-weight models purpose-built to point a developer at the right file when a known vulnerability shows up in a software dependency. Both are downloadable from Hugging Face, both run on a single GPU or on CPU at small sizes, and both are cheap enough that you can run them on every commit without thinking about it.
For anyone who self-hosts models or runs their own CI, Antares is a useful tool in the same local-AI toolkit that already covers chat, transcription, and embeddings. For anyone paying OpenAI or Anthropic to do static code analysis, the price gap on the same job is roughly two orders of magnitude. The deployment shape is the same as running any other open-weight model on your own box - the same Ollama and Open WebUI workflow covered in our self-hosting guide handles the chat layer, and Antares slots into the same inference stack with a narrower job.
What Antares actually does
Antares is not a code generator and not a chatbot. It is a “vulnerability localization” model: feed it a CVE description, an advisory, or a CWE pattern, and it returns a ranked list of source files that are most likely to contain the relevant bug, along with the terminal-style exploration trace it used to get there. That is the piece of work a security engineer does today by hand, often with grep, ripgrep, and a copy of the CVE database open in another tab. (Cisco blog)
SiliconANGLE’s coverage describes the design as starting from a vulnerability description and producing a ranked shortlist of files, while SecurityWeek adds that each model is “compact enough to run locally” and supports “local or on-premises operations.” Amin Karbasi, VP and chief scientist at Cisco Foundation AI, told SiliconANGLE the goal is “a system where all security practitioners, regardless of on-prem or resource constraints, can effectively incorporate AI in everyday security operations.”
The point of shipping the smaller models is reach. The 350M-parameter variant is the one a developer laptop or a small CI runner can actually run, and the 1B-parameter variant gives more accuracy on a single beefier GPU. Cisco says a larger 3B-parameter version is forthcoming and will be retained for Cisco’s own products rather than released externally.
The cost gap is the headline
The number that matters for a self-hoster is what Antares costs to run on a real workload. To produce a benchmark, Cisco built its own “Vulnerability Localization Benchmark,” a 500-entry set used to compare models on the same task. The numbers reported by SiliconANGLE: running Antares on the benchmark costs under $1 and finishes in about an hour. Running GLM-5.2, the strongest open-weight model Cisco compared against, costs roughly $12.50, about 15.2x more. Running GPT-5.5, the strongest closed-source model in the same comparison, costs about $141, about 172x more than Antares, and takes roughly four and a half hours. SecurityWeek reports the same 172x figure and the same 15.2x figure against GPT-5.5 and GLM-5.2 respectively.
Cisco also says Antares beat “about a dozen larger open- and closed-weight models on accuracy” in the same evaluation, per SiliconANGLE. External academic commenters cited on the Cisco launch blog frame the result as a cost-accuracy crossover. Reza Shokri, associate professor at the National University of Singapore, says Antares-1B recognizes vulnerabilities “across several weakness classes and languages” with a CLI that “packages the model’s capabilities into a practical workflow.” Amin Saberi, professor of management science and engineering at Stanford University, calls the result “near-frontier accuracy on secure code reasoning at a fraction of the cost.”
The license posture also matters. Cisco’s launch blog describes Antares as “open-weight” and SecurityWeek distinguishes that from open source: “Open weight is not quite open source, but the ‘weights’… can be controlled by the user.” Both articles confirm the weights are downloadable from Hugging Face, and the collection page lives at fdtn-ai on Hugging Face. Cisco previously released Foundation-Sec-8B, Foundry Security Spec (May 2026), and Project CodeGuard (October 2025), per SiliconANGLE, so the company has a track record of putting security AI artifacts into the open.
Why a self-hoster should care
Antares is the first open-weight model we have seen that is purpose-built for vulnerability localization rather than adapted from a general-purpose chat model. That distinction matters for two reasons. First, the model is small enough that the entire workflow can sit inside a CI pipeline: scan on every pull request, run on a single GPU, log the ranked file list to the PR comment, and never send the diff off-box. Second, the cost is low enough that a small team or a solo developer can afford to use it as a default step in development, which is the only way a security tool actually changes outcomes. Tools that cost dollars per scan get disabled; tools that cost fractions of a cent get used.
For readers who already run Ollama, vLLM, or LM Studio, the practical deployment looks like the existing local coding-model setup but with a narrower job: pull the weights, run inference against the CVE database you already maintain, and emit a short list of files rather than a paragraph of prose. Antares is not going to replace a security engineer, but it is a useful filter that runs before the engineer opens the ticket.
It also pairs naturally with the other open-weight security tooling the site has tracked. The UK AI Security Institute’s July 17 evaluation found that GLM-5.2 and DeepSeek V4-Pro are within four to seven months of closed-frontier models on cyber tasks at roughly 1% to 50% of the cost - the same AISI finding we covered in our analysis of the open-weight gap. Antares is a more focused answer to a narrower question, and the cost-per-vuln story fits the same pattern.
What This Means
Antares turns a security workflow that used to require either a paid static-analysis subscription or a chunk of an engineer’s afternoon into something a 350M-parameter model can run for fractions of a cent. The relevant comparison for a self-hoster is not “which model is most accurate” but “which model can I run on every commit without changing the bill,” and Antares is the first open-weight release we have seen where that comparison lands decisively in favor of running the model yourself.
The narrower point worth keeping: Cisco shipped two open-weight models, kept a third larger variant for itself, and published benchmark numbers that put the open variant ahead of closed-frontier alternatives on cost by two orders of magnitude. That is a deliberate business decision to seed the developer tooling ecosystem, and it tracks with the company’s earlier open releases.
The Bottom Line
Cisco’s Antares-350M and Antares-1B are open-weight, locally runnable vulnerability-localization models that cost about 172x less than GPT-5.5 on Cisco’s own 500-entry benchmark, beat about a dozen larger models on accuracy in the same test, and ship with a CLI built for CI. For self-hosters and small teams, they are the most direct local-AI answer yet to “who is going to read every CVE against our codebase.”