Cisco's open-weight security models challenge Google and OpenAI
Small but powerful models make software vulnerability detection faster and cheaper.
1Cisco's new security-focused models
Cisco has released two open-weight models, Antares-350M and Antares-1B, specifically trained to detect known security vulnerabilities in software code. Available on Hugging Face, these models are currently accessible only to vetted users, ensuring controlled deployment in the security community.
2Advantages of local processing
Unlike cloud-based LLMs, Antares models are designed to run locally, keeping proprietary code within organizational boundaries. This architecture meets strict privacy requirements and reduces the risk of malicious exploitation, making these models ideal for sensitive environments where data sovereignty is crucial.
3Performance comparison with larger models
Cisco claims Antares-1B outperforms Google's Gemini 3 Pro and matches Z.ai's GLM-5.2. The upcoming Antares-3B reportedly surpasses both GLM-5.2 and OpenAI's GPT-5.5 in vulnerability detection. These benchmarks position Cisco's SLMs as serious competitors to established AI security solutions.
4Speed and cost benefits
The compact size of Antares models enables significantly faster processing at a fraction of the cost. While scanning 500 repositories takes Antares just 15 minutes, larger models require 5 hours. The cost difference is equally striking: under $1 versus $100-$150 for frontier models, making security analysis more accessible.
5A different training approach
Cisco trained Antares models unconventionally, designing them as investigators rather than chatbots. Capable of executing multiple search strategies simultaneously, these models can pinpoint minute vulnerabilities like finding a needle in a haystack - a capability that sets them apart from general-purpose LLMs.
6A new era in security
Cisco's innovation may herald a paradigm shift in vulnerability detection. Small, focused, and locally deployable models present a compelling alternative to expensive, slow LLMs. As privacy concerns grow, such specialized solutions will likely gain importance, potentially reshaping the cybersecurity AI landscape.