Step 5 (Jieyue) has open-sourced its 600-billion-weight model, featuring a 92-layer deep architecture that reduces index overhead for millions of tokens to one-eighth.

Beating AI News Flash: StepFun has open-sourced the Step 5 Preview model weights, with FP8 and BF16 versions now available on Hugging Face and ModelScope, while supporting inference code has also been released simultaneously. The Step 5 Preview model has a total of 600 billion parameters, with only approximately 27 billion activated per generation, and supports a 1-million-token context window. A newly released technical blog reveals the model adopts a 92-layer "narrow and deep" architecture, allowing information to be processed through more layers to support multi-step reasoning in complex tasks. It also introduces Sparse Grouped Query Attention (Sparse GQA), merging indices of 8 adjacent tokens to reduce indexing and Top-k filtering overhead to 1/8 of that of per-token indexing schemes. During the reinforcement learning phase, the team incorporated context compression into training, enabling the agent to organize progress when tasks are too long before resuming execution. The model uses the custom StepFun Model License 1.0, which permits general commercial use, but hardware manufacturers such as mobile phone and automotive companies must sign a separate license agreement if integrating the model’s capabilities into commercial products. Enterprises operating MaaS (Model as a Service) businesses with total revenue exceeding $20 million over 12 consecutive months (including their own and affiliated entities) also need to obtain separate permission for commercial use.