Step 5 Preview

What it is

Step 5 Preview is a preview release of StepFun's flagship foundation model, published in September 2026 ^[1][2]. The English Wikipedia entry dates the release to 18 September, while the Chinese entry, citing StepFun's own announcement, gives 20 September ^[1][2]. The word "preview" in the name marks it as a version put out ahead of a final release. On StepFun's Open Platform the model is served under the name step-5-preview ^[4].

Architecture

Step 5 Preview is a sparse mixture-of-experts (MoE) model with 600 billion total parameters and 27 billion active parameters ^[3]. A mixture-of-experts layer contains many separate expert subnetworks, and a routing mechanism sends each token to only a few of them; that is what "sparse" refers to. The result is that the two parameter counts measure different things. The total, 600 billion here, describes how much the model as a whole holds. The active count, 27 billion, describes how much is used for any single token, and therefore roughly how much computation each token costs ^[3]. A model built this way can carry a very large total parameter count without per-token compute rising in proportion.

Inputs, outputs and context

Step 5 Preview supports a context window of up to 1 million tokens, and on StepFun's API it can produce up to 64K output tokens ^[4]. The context window is the span of material the model can take into account at once, including the prompt, images and earlier turns of a conversation; a window of this size is what makes it possible to hand the model a large body of text in a single request. It accepts text, images and video as input and returns text ^[4]. A single request can carry up to 60 images ^[4]. Images may be JPG, JPEG, PNG, WebP or static GIF, and the API offers low and high detail settings ^[4]. Video may be MP4, QuickTime or Matroska, supplied by URL, as Base64 data or through StepFun's Files API ^[4].

Positioning and reported results

StepFun positions the model as its flagship for agentic work, claiming frontier-level performance in software engineering and professional knowledge work, with particular strength in finance ^[4]. Agentic work here means tasks the model carries out over many steps, such as reading files, running commands and checking results, rather than answering a single question. The model card reports a Terminal-bench score of 33.3 and a Deep SWE score of 67.7, with results as of 20 September 2026 measured under the high reasoning setting ^[3]. These are the figures StepFun published with the release.

Weights, licence and deployment

The weights are published on Hugging Face under the StepFun Model License 1.0, described as an Apache-2.0-style licence with additional attribution and commercial-licence conditions ^[3]. The same model is served through the StepFun Open Platform as step-5-preview and can be deployed with vLLM, SGLang or Transformers ^[3][4], so it is available both as a hosted endpoint and as something a user can run on their own hardware. Usage is billed on actual input and output tokens, with rates set on StepFun's pricing page ^[4]. The charge therefore follows the number of tokens sent and generated, rather than the number of requests.