Switti

Switti is an open-source framework for training and evaluating large language models on tool use and multi-step decision-making. Developed by Yandex Research, it focuses on enabling models to plan, select, and call external tools in a structured way rather than relying on ad-hoc prompting. Switti introduces a unified interface for defining tools, their input-output schemas, and usage constraints, allowing researchers and practitioners to systematically study how models interact with APIs, retrieval systems, and other computational resources.

The framework supports supervised fine-tuning and reinforcement learning setups, making it suitable for experimenting with different training strategies for tool-augmented agents. It includes curated benchmarks and evaluation protocols that test capabilities such as tool selection, argument construction, error recovery, and long-horizon reasoning.