Antares is a family of compact language models (350M, 1B, 3B) purpose-built for agentic vulnerability localization. Given a CWE description and terminal access to a repository, Antares identifies vulnerable source files through multi-turn reasoning and tool use.
Built on Granite 4.0 and trained via SFT + GRPO with verifiable rewards, Antares-3B approaches GPT-5.5 on the 500-entry Vulnerability Localization Benchmark while outperforming all open-weight models including systems 250x larger. The full evaluation completes in ~15 minutes on a single H100 at less than $0.002 per task.
All Antares variants share a common architecture and training pipeline, differing only in parameter count. Each operates as an autonomous security agent: receiving a CWE description, exploring a repository via terminal commands, and submitting its findings.
Why small models? Security tools must run close to code. Local inference avoids supply-chain risk from sending source to third-party APIs. Multi-turn agent loops amplify latency and cost linearly with turns. Antares-3B generates >1,500 tokens/sec on a single GPU, completing the full 500-task sweep in ~15 minutes.
Evaluated on the Vulnerability Localization Benchmark: 500 real-world entries across 6 ecosystems and 147 CWE categories. All scores are means over 3 independent runs. Temperature 0.3.
Capability cliff: A clear separation exists between the top cluster (0.186-0.229 File F1) and all remaining systems. Parameter count alone does not predict performance: Llama-3.3-70B scores 0.012, Qwen3.5-122B-A10B scores 0.091, while Antares-3B (3B) scores 0.223.
Antares completes the full 500-task evaluation in ~15 minutes on a single H100 at $0.82 total ($0.00164/task). This enables continuous-integration deployment at negligible marginal cost.
Two-stage post-training transforms a general-purpose Granite checkpoint into a domain-specialized vulnerability localization agent.
SFT corpus: cybersecurity reasoning (71.5%), deep research (13.1%), code search trajectories (15.4%). Semantic conditioning on intermediate-layer representations yields 5.1x improvement at 350M scale.
Multi-component verifiable rewards in isolated Docker workspaces. Reward: File F1, submission behavior, tool compliance, exploration incentive. Infrastructure: veRL + vLLM, 8xH100, FSDP.
GRPO reduces run-to-run variance by 42-65% and reshapes agent behavior in scale-dependent ways. Smaller models adopt aggressive search-flooding strategies; the 3B model learns balanced search-and-verify patterns.
Emergent specialization: These strategy differences are not explicitly programmed. GRPO discovers scale-appropriate strategies: models with limited reasoning capacity maximize information throughput via grep flooding, while the 3B model can afford to read and verify before committing.
Antares transfers to benchmarks it was never trained on, demonstrating that vulnerability localization training produces general agentic capabilities.
No task-specific post-training: Antares receives zero CodeScout SWE-Bench or tool-use/function-calling data during post-training. The terminal navigation policy learned from vulnerability localization transfers directly to software engineering localization and structured tool calling.
Race against Antares-1B on real vulnerability localization. You both get the same repo and CWE description. Find the vulnerable file(s) before the model does.
Note: The model side replays traces recorded from live inference runs, not live weights on your browser.