StepFun Announces Step 5 Preview 600B MoE Model Set for October Open-Source
Chinese AI startup StepFun has introduced Step 5 Preview, a 600-billion parameter sparse Mixture-of-Experts (MoE) model with a 1-million token context window. The company announced plans to fully open-source the model's weights on October 15. Step 5 Preview ranks among the top three open-source models on the Artificial Analysis Intelligence Index while running at one-eighth the single-task cost of commercial models like Claude Opus 5. This offers the open-source community accessible high-level performance for complex software engineering and long-duration agentic tasks. The model uses a sparse MoE architecture that activates only 27 billion of its 600 billion total parameters during inference, while natively processing text and multimodal vision inputs. In benchmark tests, it demonstrated long-horizon autonomous capabilities, such as configuring, coding, and debugging an ESP32-S3 microcontroller hardware setup for over 3 hours continuously.
## BACKGROUND
Sparse Mixture-of-Experts (MoE) is an architecture that routes inputs to specialized sub-networks, allowing models to scale to huge total parameter sizes while keeping computational cost low by activating only a subset of parameters. The Artificial Analysis Intelligence Index is a prominent composite benchmark that evaluates AI capabilities across math, coding, reasoning, and agentic workflows.