AgentPlays
StepFun releases open weights for Step 5 Preview, a 600B agentic model with up to 1M context
StepFun published the weights of its flagship model for agentic work on Hugging Face. Step 5 Preview is a sparse mixture-of-experts model with 600B total and 27B active parameters, up to 1M tokens of context, text and image input and text output. FP8 and BF16 checkpoints are on the main model card, and an NVFP4 version is also available. The license is the StepFun Model License 1.0, Apache-style with commercial conditions. The benchmark results on the card are StepFun's own claims. StepFun's API lists it at $1 per 1M input tokens (cache miss) and $2.70 per 1M output.
How you’d use it: Run a long-context coding agent on your own GPU cluster instead of an API.