On October 15, 2026, the Chinese AI company StepFun released the weights (the trained data that makes up the model) of its flagship model Step 5 Preview. The release came exactly as announced on September 20, when the company said it would publish them on October 15, and anyone can now download the model and run it on their own systems. The license terms, which had not been disclosed at the time of the announcement, were also made public this time.
What is Step 5 Preview?
Step 5 Preview is a model aimed at "agent" use, where AI works through tasks on its own while using tools, in areas such as programming, specialized knowledge work and financial analysis. It uses an "MoE" design in which, of 600 billion (600B) parameters in total, only 27 billion (27B) are used in each pass. It can read up to 1 million tokens at once.
According to StepFun, it scored 44 points on the overall index of Artificial Analysis, a site that compares AI models. We covered the model's features, evaluations and API pricing in detail in our article at the time of the September announcement.
Released weight formats
The weights are available on Hugging Face and ModelScope, in the following three formats.
- BF16: the original-precision format
- FP8: a format that shrinks the data to reduce the memory required
- NVFP4: an even smaller format for NVIDIA's newer GPUs (Hugging Face only)
The output is text. As for input, the specification table in the model card lists text and images, while the introduction and StepFun's documentation for the API version say it also supports video, so the descriptions differ.
License terms
The license is the company's own StepFun Model License 1.0. Similar to Apache 2.0, which is widely used in open source, it allows use, modification and redistribution free of charge, and includes a patent license. The main conditions are as follows.
- If a product or service using the model has more than 100 million users in a given month, or its revenue for that month exceeds $20 million, it must display "Powered by StepFun AI"
- Selling devices with the model built in requires a separate commercial agreement regardless of revenue
- Companies running a business that offers the model to others through APIs and the like (MaaS) whose total revenue, including affiliates, exceeds $20 million over 12 consecutive months need a separate agreement for commercial use
For most companies and individuals, this means they can freely use the model in their own services and internal operations.
What you need to run it
According to the model card, the NVIDIA GPUs required vary by format as follows.
- FP8: a single server with 4 B300s or 8 H200s
- BF16: a single server with 8 B300s, or 16 H200s split across 2 servers
The model card does not list requirements for the NVFP4 format. For inference (the process of running the model to produce answers), it supports vLLM and Transformers, while SGLang supports only the FP8 format. It is not something that can run on a personal computer.
CoAI's take
Step 5 Preview was first offered via API, and its weights were released about a month later as promised. Although the license has conditions on attribution and large-scale provision, it is easy to use commercially for most users. On the same day, Meta also released its large open model Muse Spark Open, so the choice of large models that can be run on your own systems grew all at once.
However, most of the performance figures come from StepFun's own announcements, and third-party evaluations of the model as an open-weight release are still to come. If you are considering adopting it, check the license terms, then try it via the API first to see whether it suits your work.