StepFun Opens Step 5 Preview Weights for Long-Horizon Agents
StepFun has released Step 5 Preview as open weights, pairing a 600B sparse MoE model with a new harness for sustained agent work.
What happened
StepFun has released Step 5 Preview as open weights, making BF16 and FP8 versions available through Hugging Face and ModelScope. The company describes the model as a 600-billion-parameter sparse mixture-of-experts system with roughly 27 billion active parameters. It has also published a technical account of the architecture and training process, alongside Step Code 0.1.4, an open-source agent harness designed to maintain goals across long, complex tasks.
The release includes practical deployment paths rather than only a research announcement. StepFun says users can serve the model through OpenAI-compatible interfaces using vLLM or SGLang, run it offline with Transformers, and adapt it across multiple chip platforms. That combination matters because open weights become strategically useful only when developers can actually obtain, serve and modify them.
Why it matters
The most important part of the release is the alignment between the model and the harness. StepFun is presenting long-horizon work as a systems problem: model behavior, memory of the goal, tool execution and deployment hardware must be designed together. The open weights also give Chinese and international developers a new large-scale alternative at a time when access to frontier APIs remains commercially and geopolitically constrained.
The caveat is equally important. StepFun's headline parameter count does not establish that the model matches the best closed systems, and the company has not yet supplied independent evidence across a broad set of agent benchmarks. Running a 600B-total-parameter model will also remain expensive even with sparse activation. Still, this is a consequential release because it turns StepFun's agent strategy into a downloadable artifact rather than a product promise.