Cisco’s tiny open-weight AI outperforms Gemini at bug hunting

Cisco is challenging the prevailing logic of the AI arms race with a new family of compact, open-weight models designed specifically for cybersecurity. The company’s Antares-350M and Antares-1B models, released by its Foundation AI team, are small enough to run on local hardware yet powerful enough to outperform much larger systems from Google and OpenAI at one critical task: finding software vulnerabilities.

The core argument from Cisco is a direct counterpoint to the industry’s obsession with ever-larger frontier models. You do not need a massive, expensive AI to hunt for bugs in code. More importantly, you should not have to ship your proprietary source code to a cloud provider like Google or OpenAI to get the job done. Antares is designed to keep that code on your own machines.

These are not general-purpose chatbots. The Antares model works through a code repository with the methodical approach of a human investigator. It starts with a vulnerability description, searches the codebase, reads candidate files, and backtracks when a path leads nowhere. It then returns a ranked list of the files most likely to contain a flaw. The name, drawn from a red supergiant star roughly 1,000 times the size of the sun, hints at the model’s outsized ambition.

Cisco’s internal benchmarks tell a striking story. On a 500-task evaluation for vulnerability localization, the Antares-1B model beats Google’s Gemini 3 Pro and matches Z.ai’s GLM-5.2. The company’s unreleased 3-billion-parameter variant, held back from public release, reportedly tops both GLM-5.2 and OpenAI’s GPT-5.5.

The real advantage, however, is in speed and cost. Chief AI scientist Amin Karbasi told The Register that Antares can clear 500 code repositories in just 15 minutes. Frontier models require roughly five hours for the same workload. The cost difference is even more dramatic: less than $1 for Antares versus $100 to $150 for the larger systems.

The efficiency comes from a focused design. These models are trained to search, not to chat. Because each is tiny and nimble, it can run many searches simultaneously and switch tactics when a line of attack stalls. Karbasi likened it to a bicycle weaving past a truck on a busy London street.

Running locally is the other major selling point. Being small and open-weight, the models keep proprietary code entirely on a company’s own infrastructure. Cloud-based systems, by contrast, send that code to an AI provider’s servers for scanning. For firms in regulated industries, keeping everything in-house is not a luxury. It is a necessity.

There is a catch, and Cisco is honest about it. A tool that finds flaws is also one that helps attackers exploit them. As Reza Shokri, a computer scientist at the National University of Singapore, put it, AI agents now “write more of the code, and are growing capable of exploiting it.”

That is why Cisco is gating access. The models sit on Hugging Face, but the company is approving access case by case, prioritizing academic, nonprofit, and smaller public-sector security teams. The strongest 3-billion-parameter model remains unreleased. The goal, Cisco says, is to prevent advanced AI defense from becoming a luxury good, just as attackers begin wielding AI of their own. The edge, the company bets, is no longer the size of the model. It is who gets to aim it.