StepFun releases Step5Preview weights, a 600B MoE model for long agent tasks

StepFun has officially released the weights of its Step5Preview model, covering BF16 and FP8 versions, along with a technical blog and open source tools. The model uses a sparse MoE architecture with 600B total parameters, only 27B active per token, and supports a 1M token context window with native text and visual input. Step 5 Preview is built for long agent tasks and performs well in coding, software engineering, knowledge work, and finance. It recently ranked second globally on OpenRouter's weekly and daily call charts. The weights are live on Hugging Face and ModelScope, and developers can deploy it with vLLM or SGLang, or run it offline with Transformers. The technical blog explains that the model uses a 92-layer deep architecture for multi-step reasoning, Sparse GQA that cuts indexing costs to one-eighth, FP32 precision for more stable training, and reinforcement learning with context compression to handle long task runs.