Quick answer
StepFun launched Step 5 Preview on September 20, 2026, with a one-million-token context window and text, image, and video input. Hosted access and downloadable weights are separate milestones: StepFun says the model is available through its products and API, with open weights planned for October 15. This is a source-based release explainer, not a hands-on model test.
Hosted access now; open weights are still a future release
StepFun positions Step 5 Preview for coding and professional knowledge work. Its announcement says access is available through StepFun products and the API. The October 15 open-weight release is a stated plan, not a download available at this article's publication. Check the eventual model card, license, and serving requirements before planning a self-hosted deployment.
Text, image, and video input does not mean every app supports them
The official documentation lists the model ID step-5-preview, a 1M-token context window, text/image/video input, and text output. It also documents tool calling and structured output. Crucially, the integrating application supplies search, code execution, and other tools; the model does not independently gain access to a computer or external services.
Sources: StepFun
Benchmark results need their evaluation settings
StepFun publishes a mixture of public and internal benchmark results. One important footnote says its HLE-with-tools result uses a text-only subset, while several comparison models use the full dataset; those settings are not directly comparable. This article does not turn the release charts into a universal model ranking or claim to reproduce them.
Sources: StepFun
A small document exercise is more useful than a context-size contest
Before switching a real workflow, try this editorial exercise in a product that explicitly offers the model. It is a proposed check, not a recorded Step 5 Preview or Chat AI result.
- Choose two short, non-sensitive documents with different revision dates and one deliberately changed deadline. Record the correct deadline yourself.
- Ask the model to identify the newer revision, explain the conflict, and quote the supporting passage with its heading or page location.
- Ask one question the documents cannot answer. Check whether the response identifies the missing evidence instead of inventing it.
- Review the answer against the originals, then record corrections needed and the time spent reviewing. Use the same materials when comparing another model.
Frequently asked questions
What readers usually ask
Can I download Step 5 Preview weights today?
As of September 20, StepFun's announcement schedules open weights for October 15. Hosted product and API access should not be confused with released weights.
Can Step 5 Preview generate video?
Its documentation lists video as an input and text as the output. Video understanding is not the same capability as video generation.
Does a 1M-token context window guarantee a correct document summary?
No. Check important details against the source documents. The exercise above tests revision handling, evidence location, and whether the answer admits missing information; it is not a comprehensive benchmark.
Evidence