Cisco Antares downsizes AI security to 350M parameters without replacing traditional scanners

Most AI model news still moves in the same direction: more parameters, longer context windows, higher benchmark scores, and more expensive inference. Cisco Antares moves in the opposite direction. Its smallest model has just 350 million parameters, while its more capable variant reaches 1 billion. Neither model is asked to write entire applications, answer generic security questions, or perform penetration testing. They are trained for a specific task: receiving a CWE description, exploring a repository via a read-only terminal, and identifying which files most likely contain implementation flaws.

On Cisco's new benchmark, Antares-1B achieved a File F1 score of 0.209, approaching GPT-5.5 xhigh at 0.229 and outperforming several significantly larger general-purpose models. Antares-3B achieved 0.223, but weights for this top-performing variant have not been publicly released. Cisco also calculated a full 500-task evaluation sweep in approximately 15 minutes on a single H100 GPU for under one dollar, compared to over one hundred dollars for frontier APIs in their setup (S1, S2, S13).

Antares performs localization, not universal vulnerability discovery

The term vulnerability detection often combines several distinct workflows. A tool can recognize insecure patterns in a function, trace data flow, map dependencies to CVEs, execute dynamic tests, generate exploits, suggest patches, or determine whether an advisory applies to an internal repository. Antares specifically addresses file localization within that security pipeline.

Its primary inputs are a CWE identifier and a generic category description. The model does not receive full advisories, severity hints, file location hints, patches, or candidate lists. It gets read-only terminal access to the repository with a maximum budget of 15 commands. Within this loop, the model uses grep, find, cat, ls, and other standard Unix utilities to inspect patterns, read files, follow naming conventions, and revise hypotheses when initial trails lead nowhere. The output is a list of candidate file paths, not proof-of-concept exploits or verified patches (S2, S3, S4).

Search policy is the primary product

Antares is built on IBM Granite 4.0 base checkpoints, but its localization capability is not merely an inherent property of the base model. Technical evaluations show that base Granite 4.0 1B achieves near-zero localization scores, whereas post-trained Antares-1B reaches 0.209. The primary gains stem from specialized post-training policies rather than raw parameter scaling (S2, S8).

Training begins with supervised fine-tuning on a curated corpus comprising 71.5% cybersecurity reasoning, 13.1% research and general reasoning, and 15.4% code-search trajectories. This is followed by Group Relative Policy Optimization (GRPO) on multi-turn agent interaction trajectories. Rewards evaluate file-level localization precision, submission formatting, tool-use compliance, and exploration trajectories (S2, S3).

Three model sizes, but only two are publicly accessible

Cisco announced 350M, 1B, and 3B parameters in the Antares family. Antares-350M features a 32K context window for edge and lightweight deployments. Antares-1B features a 128K context window for workstations and developer hardware. Antares-3B also offers 128K context for local GPU deployments. However, the official Hugging Face collection currently hosts only the 350M and 1B weights (S2, S5).

Both released models use an Apache 2.0 license, but download access requires manual approval through a request form. Cisco enforces vetting to route model access primarily to cyber defenders while withholding the 3B model for internal security product integration (S3, S4, S13).

Conclusion

Antares illustrates how specialized domain training can deliver high utility without requiring massive parameter counts. By training 350M and 1B models on terminal interaction trajectories and verifiable search rewards, Cisco built an agent that performs repository localization at a fraction of frontier model costs.

However, benchmark results should be evaluated in context. VLoc Bench was created by the same research group and awaits independent replication. Overall task completion rates indicate room for progress, with 190 of 500 tasks remaining unsolved across all models. The primary value of Antares lies in enabling high-frequency, low-cost localization checks alongside static analysis tools and human triage workflows.