StepFun releases Step 5 Preview as open-weight 600B mixture-of-experts model

StepFun has released Step 5 Preview as open weights, publishing FP8 and BF16 checkpoints on Hugging Face and ModelScope. The move puts a large-scale open model directly in developers’ hands rather than behind an API, according to the company’s announcement covered by MarkTechPost on October 16, 2026.

Step 5 Preview is a sparse Mixture-of-Experts model, an architecture that activates only a subset of its parameters for each input. StepFun says the model has 600 billion total parameters, with 27 billion of them active per token. That design is meant to give the model a large overall capacity while keeping the computational cost of each individual request closer to that of a much smaller dense model.

The model supports context windows of up to 1 million tokens, the units a language model reads and writes text in. StepFun positions Step 5 Preview as built for long-horizon agentic work, meaning tasks that require a model to track state, tools, and intermediate steps across long stretches of interaction rather than answer a single prompt in isolation. A context window that large is relevant to that use case because it lets a model hold more of a task’s history, documents, or tool outputs in view at once.

The source material made available for this report is limited to a short summary of the release, and it does not include benchmark results, license terms, pricing, or details about how the model was trained. It also does not specify which agentic benchmarks or evaluations, if any, StepFun used to support its design claims. Those gaps mean the model’s actual performance against other open models remains unverified for now, and any such comparison should wait for StepFun’s own technical documentation or independent testing.

Where it fits

Step 5 Preview’s release adds to a run of open-weight model launches from Chinese AI labs over the past year, several of which have targeted agentic capability and long context as selling points rather than raw benchmark scores alone. By publishing both FP8 and BF16 checkpoint formats, StepFun is giving developers a choice between a lower-precision format that reduces memory and compute requirements and a higher-precision one that may preserve more accuracy, letting teams pick based on their own hardware constraints.

What happens next depends on how quickly outside developers can run and test a 600-billion-parameter model, even one with only 27 billion parameters active at a time. The hardware required to load and serve the full checkpoint will limit who can test it directly. Independent evaluations, a formal technical report from StepFun, and any benchmark comparisons against other open agentic models are the concrete markers to watch for in the weeks following this release.