This paper introduces a multi-LLM agent framework called α-UMi, which addresses the limitations of single-LLM agents in tool learning tasks. The framework decomposes the tool learning process into three components: planner, caller, and summarizer, each implemented by a separate LLM. This modular design allows for more efficient training and enables the use of smaller LLMs for specific tasks. The paper proposes a global-to-local progressive fine-tuning (GLPFT) strategy to train the multi-LLM system, which first fine-tunes a backbone LLM on the entire dataset and then fine-tunes each component on its respective sub-task. Evaluation on various tool-use benchmarks shows that α-UMi outperforms traditional single-LLM approaches, demonstrating its effectiveness in tool learning. The framework also highlights the potential of smaller LLMs in achieving competitive performance through decomposition and specialized training. The study further discusses the advantages of the multi-LLM approach, including improved flexibility, adaptability, and performance in complex tasks. The results indicate that α-UMi achieves superior performance across multiple metrics, including plan accuracy, action accuracy, and answer accuracy, while maintaining lower computational costs compared to single-LLM systems. The paper also presents case studies and analysis of data scaling laws, showing the framework's effectiveness in real-world scenarios. Overall, the research contributes to the development of more efficient and effective LLM-based agent systems for tool learning tasks.This paper introduces a multi-LLM agent framework called α-UMi, which addresses the limitations of single-LLM agents in tool learning tasks. The framework decomposes the tool learning process into three components: planner, caller, and summarizer, each implemented by a separate LLM. This modular design allows for more efficient training and enables the use of smaller LLMs for specific tasks. The paper proposes a global-to-local progressive fine-tuning (GLPFT) strategy to train the multi-LLM system, which first fine-tunes a backbone LLM on the entire dataset and then fine-tunes each component on its respective sub-task. Evaluation on various tool-use benchmarks shows that α-UMi outperforms traditional single-LLM approaches, demonstrating its effectiveness in tool learning. The framework also highlights the potential of smaller LLMs in achieving competitive performance through decomposition and specialized training. The study further discusses the advantages of the multi-LLM approach, including improved flexibility, adaptability, and performance in complex tasks. The results indicate that α-UMi achieves superior performance across multiple metrics, including plan accuracy, action accuracy, and answer accuracy, while maintaining lower computational costs compared to single-LLM systems. The paper also presents case studies and analysis of data scaling laws, showing the framework's effectiveness in real-world scenarios. Overall, the research contributes to the development of more efficient and effective LLM-based agent systems for tool learning tasks.