Step 5 Preview offers self-hosting with multi-GPU configurations

The official deployment guide lists four B300 or eight H200 graphics processors on one node for FP8 serving with vLLM or SGLang. For BF16 serving with vLLM, it lists eight B300s on one node or sixteen H200s across two nodes. These are starting configurations: available memory, the runtime and the cache holding previous computations affect usable context and simultaneous requests.

StepFun explains the design in its October 15 technical article. Long tasks with repeated tool use make processing earlier information expensive. The network has 92 processing layers and combines attention to nearby information with selective access to longer history. Its training-data workflow also divides responsibilities: experts define principles and audit critical points, agents coordinate within those principles, and reusable code performs routine operations. The company describes the system as still evolving.

The hosted service’s documentation lists text, image and video input with text output. StepFun positions it for questions spanning documents, organizing research, code changes and test creation. Crucially, the integrating application supplies search, code execution and other tools; the model itself does not access local files or external services. It supports tool calls and structured responses in JSON, a format for organized data. The service bills actual input and output token usage.

MarkTechPost summarizes the release and links the provider’s repository and technical explanation.