zxa11/qwen3-4b-router

Qwen3-4B-Router: An Agent Routing Model

The zxa11/qwen3-4b-router is a specialized 4 billion parameter model, built upon Qwen/Qwen3-4B-Instruct-2507, designed to act as a next-step routing mechanism for software engineering agents. Its core function is to analyze an agent's current trajectory and predict the characteristics of the upcoming, unexecuted step. This prediction includes the phase, phase_detail, and simple_type of the next action.

Key Capabilities & Features

• Intelligent Routing: Determines whether the next agent step can be handled by a lightweight "simple" model or requires a more robust "non_simple" model, based on predicted complexity and risk.
• Contextual Input: Processes task descriptions, early history summaries, and recent complete main round trajectories to make informed routing decisions.
• Structured Output: Generates a single-line JSON output with fixed fields: phase, phase_detail, and simple_type.
• High Accuracy: Achieves significant improvements over the base model, with SFT training boosting phase accuracy to 86.2%, phase_detail to 76.0%, and simple_type to 72.5%.
• Deterministic Operation: Designed for temperature=0 inference to ensure consistent and reliable routing decisions, crucial for agent stability.
• Customizable Thresholds: Offers an optional enhancement for threshold-based routing, allowing users to adjust the confidence level for routing to a "simple" model based on P(simple) logits.

When to Use This Model

This model is ideal for developers building sophisticated AI agents, particularly in software engineering, who need to dynamically manage computational resources. It enables efficient workflow by routing less complex tasks to faster, smaller models while reserving powerful models for critical, high-reasoning steps. Its deterministic nature and high predictive accuracy make it suitable for integrating into automated decision-making pipelines.