Cisco releases gated Antares code security models on Hugging Face
Cisco says its new Antares small language models can find known code flaws locally, with access restricted to vetted Hugging Face users.
By Dominic Okoye · Staff Writer
· 3 min read
Cisco has released two open-weight security models, Antares-350M and Antares-1B, for finding known vulnerabilities in existing codebases, with access limited to vetted users on Hugging Face. The move puts Cisco into a narrower contest with Google and OpenAI over whether code security work needs frontier-scale cloud models or smaller local systems tuned for one job.
The models are part of Cisco’s new Antares family of security small language models. DJ Sampath, Cisco’s senior vice president and general manager of AI software and platform, told The Register that the company is gating access and granting it selectively. He said Cisco is working with academic groups, nonprofits, smaller organizations and security teams at public organizations to make the tools available.
Cisco did not frame Antares as a general coding assistant. The company is positioning the models for vulnerability detection and localization: finding a specific flaw inside a codebase. Sampath said the models are intended to run locally, which means users need access to the relevant source code for scans to be useful.
That local deployment is the main commercial and security argument. Cisco said keeping analysis on an organization’s own machines means proprietary code does not have to be sent to an outside AI provider for processing. That could make the models more usable for teams with strict privacy, regulatory or internal policy limits on exposing source code to cloud services.
Cisco claims speed and cost advantages
Cisco is also making a performance claim. Amin Karbasi, Cisco vice president and chief AI scientist, told The Register that Antares-1B beats Google’s Gemini 3 Pro on Cisco’s new benchmark for measuring how efficiently AI systems identify code security flaws, and is comparable with Z.ai’s GLM-5.2. He also said a future Antares-3B model performs better than GLM-5.2 and OpenAI’s GPT-5.5 on the same test.
Those comparisons are Cisco’s claims, based on its benchmark. The 3-billion-parameter Antares model has not been released to the public, and Karbasi said Cisco will gate that model fully for communities it decides need access.
The more concrete comparison is operational. Karbasi said Antares can process a group of 500 repositories in 15 minutes, while frontier models take five hours. He also said the run costs less than $1 with Antares, compared with more than $100 to $150 using larger models.
The point is parameter efficiency. Antares-350M and Antares-1B are small by current AI standards, and Cisco argues that a model trained for search across code can run many investigations in parallel without the compute profile of general-purpose systems. Karbasi described Antares as closer to an investigator or search tool than a chatbot.
The release is also a controlled open-weight strategy rather than a fully public drop. Cisco is making two smaller models available through Hugging Face, but only after user review, and it plans tighter controls for the larger Antares-3B. That reflects the tension around AI security tooling: the same capability that helps defenders find weak code can help anyone with source access locate flaws faster.
This story draws on original reporting from The Register.
