350M and 1B dense
32K tokens
Apache 2.0
Official small-model open-weight release family from Cisco Foundation AI focused narrowly on agentic vulnerability localization rather than general chat or broad reasoning benchmarks. The public model cards describe Antares-350M and Antares-1B as decoder-only Granite 4.0 descendants trained to operate inside a terminal loop, where the model receives only a generic CWE category description and must explore a code repository with shell commands such as grep, find, and cat before submitting likely vulnerable files. Both models are trained with a two-stage pipeline of supervised fine-tuning followed by Group Relative Policy Optimization over full multi-turn agent trajectories, with rewards tied to file-level localization quality, valid terminal behavior, and exploration strategy. Cisco positions the family as compact on-premises security models for CI/CD, advisory-driven triage, and local security workflows, explicitly recommending sandboxed deployment with human oversight rather than use as a general-purpose assistant. The public benchmark tables on the VLoc Bench 500-task evaluation show the family punching above its size in this narrow domain, with Antares-1B (GRPO) reaching a File F1 of 0.209 and Antares-350M (GRPO) reaching 0.135, outperforming several much larger open and closed models on the same protocol, including Gemma-4-31B, GPT-5 Mini, Qwen3.5-122B-A10B, and Gemini 2.5 Flash. Those same tables also reference Antares-3B SFT and GRPO variants that outperform the released pair, but I could not verify a public weight release for the 3B checkpoint. The 350M card further notes that the smallest model can act as a speculative draft model for larger Antares variants, while both cards emphasize local deployment, a 15-command terminal budget, and English-only static training data with an April 10, 2025 cutoff.