StepFun releases Step 5 Preview as an open-weights 600B MoE model

StepFun, a Chinese artificial intelligence startup, has released its Step 5 Preview language model as open weights. As of October 17, 2026, checkpoints in both FP8 and BF16 precision are available for download on Hugging Face and ModelScope.

The model uses a sparse Mixture-of-Experts (MoE) design. It contains 600 billion parameters in total, but only 27 billion are activated for any given token. Step 5 Preview also supports a context window of 1 million tokens.

Built for long-horizon agents

StepFun describes the model as purpose-built for long-horizon agentic work. Such tasks require an AI system to call tools, keep track of intermediate results and stay coherent across lengthy, multi-step workflows. The large context window is meant to let the model hold extensive documents or long interaction histories in view at once.

In an MoE architecture, only a subset of specialised sub-networks runs for each input. That allows a model with very large total capacity to keep inference compute closer to that of a much smaller dense model. Shipping an FP8 checkpoint alongside BF16 also lowers the memory needed to serve it.

Another entry in the open-model race

The "Preview" label signals that this is not a final release. The available information does not include a timeline for a full version or detailed benchmark comparisons, so claims about how the model stacks up against rivals will need independent evaluation.

Shanghai-based StepFun is one of several Chinese startups building large language models. Releasing frontier-scale models with open weights has become a common strategy among Chinese AI developers, giving researchers and companies the option to run and adapt the systems on their own infrastructure rather than relying solely on hosted APIs.