New Mouse-1

Mouse-1 is a model specifically designed to assist AI agents in making “small decisions.” Whenever an agent completes a task, it often needs to make hundreds of minor judgments along the way—for example, deciding which tool to use next or whether to attempt the same operation again. Entrusting all such decisions to larger models would be both time-consuming and costly; that’s where Mouse-1 comes in—to handle these tasks efficiently.

It receives the current state of the agent and available options, and returns a probability that allows the framework to make decisions directly based on this information. Cutting-edge large models are only responsible for those tasks that truly require reasoning and code implementation. In the official demonstration, it completes the task before the large model generates even its first token; each decision it makes costs approximately 0.00004 US dollars.

It is provided in the form of an API and can be easily integrated into agent frameworks. This option is ideal for developers who wish to build their own agents and are concerned about speed and cost.