StepFun: the Shanghai startup behind the Step models

A young and ambitious startup

StepFun was founded in 2023 in Shanghai. It is one of the young Chinese companies that build their own large language models rather than relying on someone else’s. Its Chinese name combines the ideas of a “leap” and the “stars”; that notion of moving up step by step shows in the name of its models, Step.

Models built for agents

StepFun’s signature is its focus on AI agents: models that can chain actions, call tools and carry a long task through to the end. Recent versions make this explicit:

• Step-3.5-Flash: a fast, cost-efficient model built on a mixture-of-experts design that activates only a small part of the model at each step;
• Step-3.7-Flash: its successor, presented as an efficient model for real-world agents;
• Step-5-Preview: the flagship, positioned for agentic work.

Beyond text, StepFun covers voice (Step-Audio), video (Step-Video), image editing (Step1X-Edit) and even graphical-interface agents with GELab, designed to operate apps the way a user would.

A largely open policy

StepFun releases most of its models as open weights on Hugging Face and ModelScope. The Flash models use the Apache 2.0 license, which allows free commercial use. The flagship Step-5-Preview can be downloaded too, but under an in-house license with extra conditions. In practice, check the license model by model before any commercial project.

The large Step models need servers with high-end graphics cards. For a first try, the API is the easiest route.

How to access the models

StepFun offers an API on its international platform, plus a separate platform for China. The models are also available through OpenRouter, and the weights can be deployed with vLLM, SGLang or Transformers.

To see how Step compares with other open models, visit our open-source models page or the model comparison tool. The profile of MiniMax, another young Shanghai lab, makes for a useful comparison.