StepFun opens Step 5 Preview and sets open weights for Oct. 15

StepFun has made Step 5 Preview, a 600-billion-parameter model it calls its new flagship for agentic work, available through its API. The company says the open weights follow next week. Open weights are the trained model files, released for download.

StepFun announced the model Sept. 20, 2026, in a post on its official X account headlined “Advancing the Pareto Frontier.” The company called Step 5 Preview “our new flagship model for agentic work, delivering frontier-level performance across software engineering and professional knowledge work, with particular strength in finance.” The post also says “Open weights on Oct 15.” The weights are not public as of Oct. 10, 2026. StepFun’s organization page on Hugging Face, checked that day, lists no Step 5 model.

Step 5 Preview activates 27 billion of its 600 billion parameters

StepFun’s post lists the model as a mixture-of-experts design with 600 billion total and 27 billion active parameters, a 1 million token context window and vision. In a mixture-of-experts model, only part of the network runs for each token. StepFun said the model offers “substantially lower task cost at comparable intelligence” and “broad software engineering capabilities with sustained execution over long horizons.”

The API is live and OpenRouter quotes a price

StepFun’s developer platform, checked Oct. 10, shows Step 5 Preview on its front page with “Get API Key” and “Read Docs” buttons. It describes the model as “A new-generation frontier model built for real-world tasks.”

OpenRouter lists “StepFun: Step 5 Preview” as released on its service Oct. 8, with StepFun as the only provider. Prices as of Oct. 10 are $1 per million input tokens and $2.70 per million output tokens, with cache reads at $0.05 per million. The listing shows a 1 million token context, up to 64,000 completion tokens, and text, image and video input with text output. Reasoning is mandatory, with high, medium and low effort levels and medium as the default. OpenRouter describes the model as built for extended tasks across large codebases and documents, using tools and refining results over multiple steps.

We think the weights will let outsiders test the claims

We think the sparse design is how StepFun keeps its price low. A model that holds 600 billion parameters but uses 27 billion for each token leaves most of the network idle. We expect the cost per task to follow the smaller number.

The weights due Oct. 15 let outside developers check the size claim and run the model themselves. Until then, “comparable intelligence” is StepFun’s own description. We have not tested it.