StepFun Unveils Step 5 Preview for Long-Horizon Agents
StepFun’s new 600-billion-parameter sparse model targets software, research and finance agents with a one-million-token context window.
StepFun has introduced Step 5 Preview, its new flagship model for agentic work, in a release timed for the Asian daytime cycle. The model is described as a 600-billion-parameter sparse mixture-of-experts system, with roughly 27 billion parameters activated per token, a one-million-token context window and native image input.
Built for extended professional tasks
StepFun positions the preview around work that requires more than a single answer: software engineering, large-scale research, structured analysis and interactive reporting. Finance is a particular focus, with the company highlighting source verification, valuation work and the ability to sustain multi-step investigations.
The model is available through StepFun’s preview interface and documentation, but the launch needs to be read carefully. This is an API and preview-model release, not yet a completed open-weight release. StepFun’s materials indicate that model weights are planned for a later date. That distinction matters for teams deciding whether to build around the model or merely test it.
Why it matters
Step 5 Preview arrives as Chinese AI companies increasingly compete on agent reliability and work completion rather than headline benchmark scores alone. Its combination of a large context window, visual input and long-horizon execution is aimed directly at the same professional workflows being targeted by leading US and European labs.
The immediate significance is strategic: StepFun is trying to make an Asian-developed model part of the global agent stack before its weights become available. The unresolved questions are equally important—how the model performs on independent evaluations, how stable its preview capacity is, and whether the eventual open release will carry a permissive license and practical deployment requirements.