Throwing a massive, multi-billion-parameter cloud language model at your source code to find security bugs is an expensive mistake. Frontier models gobble tokens, take hours to process large repositories, and force security engineers to upload proprietary code to third-party cloud servers.
Cisco is taking a sharp turn in the opposite direction. With the release of its Antares family of small language models (SLMs)—specifically the open-weight Antares-350M and Antares-1B—the enterprise networking giant is betting that specialized, local models will outperform generalized cloud beasts at vulnerability detection.
The Real Economics of Scanning Code
When security teams run routine code audits across enterprise systems, the bill for cloud-hosted AI adds up fast. Scanning a suite of 500 software repositories with general-purpose frontier models can take up to five hours and cost between $100 and $150 per run. For a dev team pushing code continuously, those numbers quickly turn automated security into an unsustainable budget line item.
Cisco’s benchmark testing paints a very different picture for specialized small models. Running the same 500-repository scan using Antares-1B takes just 15 minutes and costs under a single dollar. That is a 20x speedup and a near-total collapse in compute expenses.
Why the massive discrepancy? Frontier models are built to generate broad semantic reasoning and creative text. They carry immense parameter weight that eats up GPUs even when doing basic pattern matching. Finding a security flaw isn't a creative writing exercise—it's a needle-in-a-haystack search across millions of lines of code where only a single line or file might be broken. Paying for frontier-model inference to scan static files is like hiring an expensive legal team to proofread routine receipts.
Local Execution Keeps Secrets on Metal
Cost and latency are only half the battle. The bigger roadblock for enterprise security teams has always been data sovereignty and supply chain risk. Uploading thousands of lines of uncompiled internal code to a third-party AI provider’s servers creates immediate red flags for compliance officers.
Because Antares-350M and Antares-1B are lightweight SLMs, they are engineered to run directly on local infrastructure. Your code never leaves your own hardware.
As DJ Sampath, Cisco’s senior vice president and general manager of AI software and platform, put it, local execution flips the security threat model. If an attacker gains access to the machine running local scans, they already hold the keys to the system anyway. Scanning locally removes the external cloud provider from your attack surface entirely. Sensitive intellectual property stays pinned to local RAM and disk, satisfying strict air-gapped or regulated compliance environments where cloud LLMs cannot legally go. This local isolation strategy fits into broader defense frameworks for securing agentic infrastructure against AI cybersecurity threats.
Why Smaller Models Excel at Search
The core reason Cisco's small models hold their own against massive competitors boils down to architecture and training intent. Antares isn't a conversational chatbot. You won't ask it to write poetry or summarize a quarterly report.
According to Amin Karbasi, Cisco's VP and chief AI scientist, Antares was trained specifically as an investigator and search engine. Finding bugs requires an iterative, adaptable search strategy. If one path of code exploration hits a dead end, the model must pivot immediately and try a different forensic tactic.
Because Antares carries a minimal parameter footprint, it can execute multiple search pathways simultaneously without swamping hardware memory. Karbasi compares the dynamic to traffic: on a jammed city street, a nimble bicycle weaves through gridlock far faster than a massive semi-truck or a private jet. Sampath summed it up even more bluntly: "Sometimes you don't need a private jet to go to a corner store." Specialized small models move fast precisely because they don't carry the dead weight of general knowledge.
Benchmark Results and Gated Access
Despite their compact size, the Antares models punch well above their weight class on vulnerability localization benchmarks.
According to Cisco's internal evaluation data, Antares-1B outperforms Google’s Gemini 3 Pro and matches the detection capability of Z.ai's GLM-5.2 in identifying known code flaws. Cisco is also developing an unreleased 3-billion-parameter variant, Antares-3B, which early tests show beating both GLM-5.2 and OpenAI’s GPT-5.5 at code vulnerability detection.
However, Cisco isn't opening the floodgates to everyone. Both Antares-350M and Antares-1B are hosted on Hugging Face under gated access. Cisco is deliberately vetting users, prioritizing academic researchers, non-profit institutions, and defensive security teams at public and smaller organizations. The upcoming Antares-3B model will remain completely gated.
The reasoning is straightforward: automated vulnerability hunters are dual-use technology. A tool that pinpoints bugs in seconds for a defender can just as easily help an adversary uncover zero-day flaws before a patch exists. Gating access strikes a deliberate balance between empowering open-source defense and preventing weaponized exploitation.
Pragmatic Security Over Flashy Demos
Cisco's strategy with Antares highlights a broader shift taking hold across enterprise IT infrastructure. For two years, tech companies competed almost exclusively on parameter scale and raw general intelligence benchmarks. But in operational security, giant cloud models are frequently impractical.
By focusing on narrow, high-impact tasks like vulnerability scanning and localization, Cisco proves that small, focused models running on local hardware can deliver better performance at a fraction of the cost. Enterprise defenders don't need giant, talkative chatbots in their security operations centers—they need fast, reliable tools that find bugs before attackers do. This pragmatic shift reflects growing demand for realistic frameworks to evaluate AI SOC tools based on actual operational utility rather than marketing claims.
Detailed coverage of Cisco's open-weight release and benchmark claims can be explored further in The Register's initial report.
