The AI industry has spent much of the past few years making models bigger. Cisco is taking the opposite approach with Antares, a new family of small language models designed specifically to find security vulnerabilities in software code.
The idea is fairly practical. Instead of asking companies to upload proprietary source code to a cloud-based general-purpose AI model, Antares is designed to run inside an organisation’s own environment. That changes the trade-off for security teams that want the speed of AI-assisted code analysis but can’t easily justify sending sensitive software outside their own infrastructure.
Cisco is initially releasing two models, Antares-350M and Antares-1B, as open-weight models. Their relatively small size is central to the proposition rather than a limitation Cisco is trying to work around. Smaller models require less compute, can be deployed locally and may be easier for organisations to integrate into existing security workflows without building expensive AI infrastructure around them. Cisco says both models are available to the broader developer and security community.
That makes Antares interesting for a reason that has relatively little to do with model size on its own. AI adoption inside companies is increasingly running into the awkward reality that some of the most valuable data is also the data organisations are least willing to hand over to someone else’s cloud.
Keeping the code inside the building
Source code sits near the top of that list. It can reveal intellectual property, internal architecture and, potentially, weaknesses that an attacker would be delighted to find. For security-conscious organisations, the convenience of cloud AI therefore comes with a fairly obvious question about what has to leave the organisation in exchange for that convenience.
Cisco’s pitch is that Antares removes part of that tension by keeping code analysis on-premises. The company specifically points to public-sector organisations, universities and businesses subject to strict data-sovereignty requirements as potential beneficiaries.
There’s a familiar idea here for anyone following the enterprise infrastructure market in South Africa. Data sovereignty has moved steadily closer to the centre of technology procurement, particularly in government and highly regulated industries. Cisco itself has already been pursuing that market through its Sovereign Critical Infrastructure portfolio in South Africa, which can be deployed on-premises, air-gapped or as part of a hybrid environment. Antares applies much the same logic to AI: sometimes the problem isn’t whether the technology can do something, but where it has to run before an organisation is comfortable allowing it to do so.
Fady Younes, Cisco’s Managing Director for Cybersecurity in the Middle East and Africa, argues that securing increasingly complex software without compromising privacy has become more urgent as organisations expand their digital capabilities. With Antares, Cisco’s answer is essentially to move the AI closer to the code rather than moving the code closer to the AI.
There’s also a broader architectural shift hiding inside that approach. Enterprise AI has often been discussed as though companies will eventually connect everything they do to increasingly capable general-purpose models. Antares represents a more specialised alternative: smaller models, trained for narrower jobs, running where the organisation needs them.
For some enterprise workloads, that could be considerably more useful than simply having access to the most capable general-purpose model available.
A model that investigates rather than merely scans
Cisco says Antares doesn’t work like a conventional rules-based vulnerability scanner. The models are designed to read a vulnerability description, search through relevant parts of a codebase and change direction when a particular line of investigation stops producing useful results. From there, they narrow down the file paths most likely to contain the vulnerability.
That approach speaks to a problem that’s becoming increasingly important as AI enters software development itself. Finding vulnerabilities isn’t necessarily the end of the job. Security teams still need to understand where the weakness sits, how serious it is, what depends on the affected code and what should happen next.
It’s a problem we’ve already seen elsewhere in AI-assisted cybersecurity. When TrendAI integrated Claude Opus 4.7 into its vulnerability research platform, the interesting problem wasn’t simply whether AI could discover more vulnerabilities. It was what happens when automated systems begin finding problems faster than security teams can prioritise and remediate them. Antares attacks an earlier part of that same workflow by trying to narrow the search before a human investigator gets involved.
Cisco says benchmark testing shows the Antares models outperform a number of larger open and closed models on vulnerability localisation while operating at substantially lower cost. Its technical announcement also makes a more specific claim about why small models are useful here: because they can run locally and quickly enough to become part of continuous software-development workflows rather than being invoked only after the fact.
Those performance claims will still need to stand up outside Cisco’s own testing, particularly once developers start applying the models to different programming languages, codebases and vulnerability classes. A benchmark showing that a small specialised model beats a larger general-purpose one is interesting. It doesn’t automatically tell us how often it will find the right file inside years of poorly documented production code.
That uncertainty doesn’t undermine the idea behind Antares, though. It makes the open-weight release more consequential because independent developers and security researchers can now test where the models work, where they fail and whether their efficiency advantage survives contact with less controlled environments.
Smaller AI may make more sense for cybersecurity
Much of the current AI market has trained companies to think about models hierarchically. More parameters suggest greater capability, and smaller models are often treated as compromises for situations where hardware, latency or cost becomes a constraint.
Antares turns that assumption around. Vulnerability localisation is a specialised problem, so Cisco is betting that a specialised model doesn’t need to be particularly good at writing poetry, answering general-knowledge questions or doing everything else expected from a frontier chatbot.
That’s a useful distinction because enterprise AI economics become strange very quickly when every problem is handed to a large general-purpose model. If a 350-million or one-billion-parameter model can reliably perform a specific security task locally, paying the compute and token costs associated with a much larger model begins to look less like ambition and more like waste.
Cisco also argues that this makes AI-assisted security more accessible to smaller teams, universities and public-sector organisations that may not have the resources to use expensive proprietary models at scale.
There are limits to how far that democratisation argument should be taken. A freely available model doesn’t remove the need for skilled security professionals, secure deployment, compute resources or people capable of deciding whether what the model has found represents a meaningful vulnerability.
AI can reduce the amount of code humans need to inspect. It doesn’t make the judgement around that code disappear.
Cisco has already been experimenting with how far that division of labour can move. At Cisco Live 2026, the company said it had used automated security pipelines to scan 1.8 billion lines of code in eight weeks, part of a broader push towards allowing AI systems to perform more of the investigation while humans concentrate on oversight and risk. As we explored in our reporting from Cisco Live, that creates its own uncomfortable question: humans can remain nominally in control while the systems they supervise begin operating at a speed and scale that humans could never reproduce themselves.
Antares fits rather neatly into that emerging security model. A small local model can continuously narrow down potential vulnerabilities, while human security teams spend more of their time deciding what those findings mean and what deserves attention.
The South African relevance is less abstract than it sounds
For South African organisations, the local-processing argument may ultimately be more persuasive than the model-performance one.
Banks, government departments, universities, healthcare organisations and large enterprises don’t all operate under identical regulatory or security requirements, but many face some version of the same problem: giving AI access to sensitive organisational information creates another set of questions about where that information travels, who has access to it and what happens to it once the model has processed it.
Antares doesn’t resolve every governance problem created by AI, and running a model on-premises doesn’t automatically make either the model or its output trustworthy. It does, however, remove one uncomfortable part of the equation by making it possible to analyse source code without first sending that source code to a third-party cloud model.
That distinction becomes more relevant in a South African security environment that’s already under pressure. Cisco’s own previous research found that only 5% of South African organisations had reached what it classified as a mature level of cybersecurity readiness, while skills shortages and the rapid adoption of AI were adding further pressure. The uncomfortable part of those numbers is that more sophisticated defensive technology is arriving in organisations that don’t necessarily have equally sophisticated security capacity.
That creates a more nuanced case for Antares than simply saying smaller AI is cheaper. Local, specialised models could lower some of the barriers to using AI for defensive security, particularly where privacy and infrastructure costs have made large cloud models difficult to justify. They can’t compensate for weak security processes, poor patch management or a shortage of people who know what to do with the vulnerabilities they find.
The most interesting thing about Antares may therefore be what it says about the next stage of enterprise AI. The industry has spent years treating model size as a proxy for progress, but businesses don’t necessarily need one enormous intelligence sitting above every problem. They need systems that can do particular jobs reliably, economically and without creating an entirely new data-governance problem in the process.
If Cisco’s models can prove themselves outside its benchmarks, Antares makes a fairly persuasive argument that some of the most useful enterprise AI may end up being the AI you barely notice: small enough to run where the work already lives, specialised enough to know what it’s looking for, and cheap enough to keep looking.