StepFun: Step 5 Preview

StepFun: Step 5 Preview is a text generation model from StepFun, released by the provider on 2026-10-08. It is described as StepFun's flagship model for agentic work, built on a sparse Mixture-of-Experts architecture with 27B active parameters out of 600B total. The model sits in the text category and is aimed at demanding, tool-driven workloads rather than simple chat alone.

The model accepts text, image and video as input, and produces text as output. Its context window is 1,000,000 tokens, and it can generate up to 64,000 output tokens in a single response. That combination suits tasks where large amounts of source material must be considered at once, whether that material is written documents, screenshots, or video footage, with a long written answer or artifact at the end.

Because Step 5 Preview performs strongly in software engineering and professional tasks, per StepFun's own description, natural uses include codebase analysis over many files, repository-level refactoring and review, and agentic workflows where the model calls tools across multiple steps. The image and video input support extends this to use cases such as inspecting UI screenshots alongside code, reviewing recorded product flows, or extracting structured information from visual material. The large context window also makes it a reasonable fit for long-document analysis, research synthesis, and multi-hour transcript or log review where the whole corpus needs to stay in view.

Through api14's OpenAI-compatible API, you call the model with the id "stepfun/step-5-preview". The supported request parameters are frequency_penalty, include_reasoning, logprobs, max_tokens, reasoning, reasoning_effort, response_format, stop, structured_outputs, temperature, tools, top_logprobs and top_p. In practice this covers sampling control (temperature, top_p, frequency_penalty, stop), output length (max_tokens), tool calling for agentic loops, structured and JSON-style output (response_format, structured_outputs), and reasoning controls (reasoning, reasoning_effort, include_reasoning) with log probability inspection via logprobs and top_logprobs. Developers already using the OpenAI SDK or client can point it at api14 and switch the model id without changing their request shapes.

Worth noting is that the model outputs text only, so tasks requiring generated images, audio or video fall outside its scope, and the "preview" label on the model id suggests treating it as a moving target rather than a frozen release. If you need a smaller, cheaper or faster model for simple classification or short chat, a lighter text model on api14 may fit better, while Step 5 Preview is best reserved for the long-context, tool-heavy and multimodal-input jobs it was built for.