StepFun Unveils Open-Source Step 5 Preview Model, Top 3 in Global AI Index

On September 20, 2026, StepFun released Step 5 Preview, a new open-source flagship foundation model. The model features a sparse MoE architecture with 600 billion total parameters and only 27 billion activated parameters. It natively supports text and vision inputs, offers a 1 million-token context window, and targets AI coding, software engineering, professional knowledge work, and finance.

Step 5 Preview scored 44 on the Artificial Analysis Intelligence Index, placing it among the top three open-source models globally and on the Pareto frontier balancing model intelligence and per-task cost. In internal finance evaluations, including FrontierFinance with 220 expert questions and 11,543 evaluation criteria, the model performed close to Claude Opus 5 on information retrieval, company valuation, and financial research tasks.

In hands-on tests, Step 5 Preview handled complex tasks across web game development, 3D asset generation, deep research, and data analysis, demonstrating the ability to break down multi-step instructions, call tools, and continuously follow constraints. It also maintained multi-condition consistency in long-context travel planning tasks involving budget, physical limits, and dietary restrictions.

Over the past quarter, StepFun also launched the Step Edge on-device model and the StepAudio 3 voice series. With Step 5 Preview, the company has expanded its lineup to cover cloud flagship, high-efficiency Flash, on-device, and multimodal models. StepFun said its models have been installed on more than 42 million smartphones, serving nearly 20 million requests per day and partnering with over 50% of major domestic phone brands.