Step 5 Preview

Step 5 Preview is a large-scale foundation model developed by StepFun, designed to support complex, long-horizon agentic workflows. It utilizes a sparse Mixture-of-Experts (MoE) architecture with approximately 600 billion total parameters, of which 27 billion are active during any single inference step. This architecture is specifically engineered to achieve high-level reasoning performance while maintaining a lower computational cost compared to dense models of similar total scale.

The model features a 1-million-token context window, allowing it to process massive datasets, such as entire software repositories or extensive financial reports, in a single prompt. It is natively multimodal, supporting text, image, and video inputs. Structurally, Step 5 Preview employs a "narrow-deep" transformer configuration consisting of 92 layers, which the creators state provides more robust paths for the multi-hop reasoning required in autonomous agent tasks.

Optimized for professional knowledge work and software engineering, Step 5 Preview includes specialized features such as tool calling, JSON mode for structured outputs, and prompt caching. Users can toggle between different reasoning effort levels—low, medium, and high—to balance response speed against analytical depth. In internal evaluations, the model has demonstrated the capacity to manage nearly 1,000 sequential tool calls within a single agentic session, emphasizing its suitability for complex, multi-step problem solving.