The model card was last updated on 15 October 2026. The release follows earlier API listings on OpenRouter and Vercel; the newly reported event is the availability of weights.
Step 5 Preview is a sparse mixture-of-experts model with 600 billion total parameters and 27 billion active parameters. StepFun lists up to one million tokens of context and up to one million output tokens. Its specification table lists text and image inputs and text output, while the card introduction also says the model accepts video. BF16 checkpoints are available. Example vLLM deployments include FP8 on one four-GPU B300 node or one eight-GPU H200 node.
The weights use StepFun Model License 1.0, not Apache 2.0. The licence adds attribution and commercial conditions: products exceeding 100 million monthly active users or US$20 million in monthly revenue must display “Powered by StepFun AI.” Hardware makers and certain large model-as-a-service businesses require separate commercial licences. The source reports StepFun’s benchmark charts but does not independently validate their results or establish real-world performance.