aiqarus-agent-4b: An Enterprise Agent Model
This 4 billion parameter model, developed by zeon01, is a fine-tuned version of Qwen3-4B-Instruct, optimized for complex enterprise AI agent workflows. It addresses critical agent capabilities such as tool-calling, multi-step planning, risk escalation, confidence calibration, and multi-agent handoff.
Key Capabilities & Improvements
- Enhanced Agent Orchestration: Designed for tool routing and task decomposition in multi-system environments.
- Improved Reasoning: Achieved a significant increase in reasoning quality (from 1.9/5 to 3.1/5) and reasoning depth (+0.93 over base model) through iterative training.
- Adversarial Awareness: Demonstrates strong injection detection capabilities (+1.83 over base model) due to the inclusion of adversarial data during training.
- Balanced Action Types: Training methodology shifted from 80% tool-calling bias to a 50/50 balance between tool and non-tool actions, improving overall performance.
- Iterative Development: The model underwent two rounds of fine-tuning, with LLM-as-judge evaluation guiding data and methodological changes, leading to recovery of action accuracy and deeper reasoning compared to the base model.
Intended Use Cases
- Enterprise agent orchestration and task decomposition.
- Multi-system workflows requiring handoff and delegation.
- Research and experimentation with small agent models.
Note: This is a research checkpoint (Round 2, SFT-only) and is not recommended for production use without further fine-tuning and alignment due to its SFT-only nature and potential overconfidence in ambiguous tool-calling cases.