StepFun Launches Step 5 Preview: A 600B-Total, 27B-Active MoE Model With 1M Context ...

StepFun has launched Step 5 Preview, a 600 billion parameter Mixture-of-Experts model activating 27 billion parameters per token with a 1 million token context window, targeting software engineering, professional knowledge work, and finance workloads at a substantially lower task cost. The model supports text, images, and video input, offers multiple reasoning modes, streaming, tool calling, and will release open weights for self-hosting on October 15, 2026.

This Wire brief sits within Fusion42's coverage of AI Frontier Models and AI Infrastructure, and 1 source has reported it.