StepFun has released Step 5 Preview as open weights, making both FP8 and BF16 checkpoints available on Hugging Face and ModelScope. The model is a sparse Mixture-of-Experts (MoE) system built specifically for long-horizon agentic work, where models need to maintain coherence and reasoning over extended interactions.
The architecture uses 600B total parameters but activates only 27B per token, a design that balances capacity with inference efficiency. It also supports a context window of up to 1M tokens, enabling it to process very long documents, conversations, or multi-step task traces without losing track of earlier context.
By releasing the weights openly, StepFun lets researchers and developers fine-tune and deploy the model on their own infrastructure rather than relying on a hosted API. The combination of sparse activation and a large context window could make Step 5 Preview a practical option for agentic systems that need deep reasoning over long horizons.