StepFun releases Step 5 Preview open weights

StepFun released FP8 and BF16 weights for Step 5 Preview, as detailed on Hugging Face. The mixture of experts model has 600 billion total parameters, activates 27 billion parameters per token and supports a context window of up to 1 million tokens.

StepFun designed the model for software engineering, tool use, document analysis and finance tasks. It accepts text and images, produces text and can work across large code repositories and long documents.

The FP8 checkpoint can run on four Nvidia B300 GPUs or eight H200 GPUs using vLLM or SGLang. BF16 deployment requires eight B300 GPUs or 16 H200 GPUs. Both serving frameworks expose the model through an OpenAI compatible API, while a Transformers implementation supports offline generation.