ToolOmni-Qwen3-4B is a 4 billion parameter causal language model developed by Shouzheng Huang et al. based on Qwen/Qwen3-4B-Instruct, specifically designed for open-world tool use. It is trained to proactively retrieve relevant tools and generate grounded multi-step tool calls, leveraging an agentic learning framework. This model excels at research on tool-use agents and benchmarking open-world tool retrieval and grounded execution, with a context length of 32768 tokens.
Loading preview...
ToolOmni-Qwen3-4B: Open-World Tool Use Model
ToolOmni-Qwen3-4B is a 4 billion parameter causal language model built upon Qwen/Qwen3-4B-Instruct, developed by Shouzheng Huang et al. for the ACL 2026 Main Conference. Its core innovation lies in enabling open-world tool use through an agentic learning framework.
Key Capabilities & Features
- Proactive Tool Retrieval: The model is trained to identify and retrieve relevant tools autonomously.
- Grounded Execution: It generates multi-step tool calls that are grounded in the task context.
- Agentic Learning: Utilizes a framework incorporating proactive retrieval, grounded execution, and reinforcement learning for complex multi-step tool-use behaviors.
- Evaluation on ToolBench-style Benchmarks: Assessed in both
with-api-list(golden-tool) and open-domain settings without predefined tool lists.
Intended Use Cases
This model is particularly suited for:
- Research on Tool-Use Agents: Ideal for exploring and developing advanced AI agents capable of interacting with external tools.
- Benchmarking: Useful for evaluating and comparing performance in open-world tool retrieval and grounded execution scenarios.
- Studying Advanced Training Paradigms: Supports research into retrieval-augmented and execution-aware training methodologies.
- Reproducing Evaluation Pipelines: Designed to work in conjunction with the ToolOmni codebase, retriever, and execution environment for consistent research.
For detailed evaluation protocols and benchmark results, refer to the project repository and the associated paper.