Cisco's small open-weight AI detects bugs, outperforming Gemini.

      Cisco's latest Antares models are compact enough to operate on your own hardware. They assert that these models outperform Google's Gemini and competing GPT-class systems in identifying bugs. They are being released as open-weight, although Cisco is selectively granting access.

      Cisco’s approach subtly challenges the race for frontier models. Their message is that you don’t need an enormous, expensive AI to discover software bugs, nor should you need to send your source code to OpenAI or Google for this purpose.

      The Foundation AI team at Cisco has launched two open-weight security models, Antares-350M and Antares-1B, designed to identify known vulnerabilities hidden within existing codebases. They are available on Hugging Face for community use, but Cisco approves access on a case-by-case basis. A larger 3-billion-parameter model is being withheld from full public release.

      These models are not intended for chatting. Antares inspects a repository similarly to a human investigator. It starts with a vulnerability description, explores the code, and examines relevant files. When a path leads to a dead end, it retraces its steps and compiles a ranked list of the files most likely to conceal a flaw. The name "Antares" is inspired by the red supergiant star, approximately 1,000 times larger than the sun.

      Compact yet powerful

      Cisco asserts that these small models significantly outperformed expectations. On their own 500-task benchmark for vulnerability localization, Antares-1B surpasses Google’s Gemini 3 Pro and matches Z.ai’s GLM-5.2. They claim the unreleased 3B model outperforms both GLM-5.2 and OpenAI’s GPT-5.5.

      The true benefit lies in speed and cost efficiency. Chief AI scientist Amin Karbasi informed The Register that Antares can process 500 code repositories in 15 minutes, as opposed to the five hours required by larger models. This translates to under $1, compared to $100 to $150 for the more substantial systems.

      The key is that Cisco developed these models for searching rather than chatting. Due to their small and agile nature, they can perform multiple searches simultaneously and adapt strategies when one approach proves ineffective. Karbasi compared this capability to a bicycle maneuvering past a truck on a busy London street.

      Keeping the code on-site

      Another attractive feature is the ability to run the models locally. Being small and open-weight, they allow businesses to keep proprietary code on their own systems. In contrast, cloud-based solutions require sending the code to an AI provider's servers for analysis. For companies in regulated sectors, in-house processing is essential.

      The downside

      Cisco's caution has its reasons. A tool designed to uncover flaws can also be used by attackers to exploit them. As Reza Shokri, a computer scientist at the National University of Singapore, observed, AI agents now "write more of the code and are becoming capable of exploiting it."

      Thus, Cisco limits access to academic, nonprofit, and smaller public-sector security teams while retaining its most advanced model. The intention is to prevent advanced AI defense solutions from becoming a luxury, especially as attackers begin to utilize AI themselves. Cisco believes that the edge lies not in the model's size but in who controls its application.