On September 20, StepFun released Step 5 Preview, a flagship model aimed at real-world agentic tasks. It uses a sparse MoE architecture with 600B total parameters and 27B activated per token, supports a 1-million-token context and vision input, and targets long-horizon tasks in coding, software engineering, research and finance.
The API is already open, and StepFun plans to release model weights on October 15. Third-party benchmark Artificial Analysis scored it around 44, on par with the 2.8-trillion-parameter Kimi K3, placing it among the global top three open-weight models. Pricing is $1 per million input tokens and $2.7 per million output tokens at 100 tokens/s.
The rapid domestic release cadence—Step, Qwen, Zhipu, Kimi—reflects an intensifying second-half race where price-performance and speed, not just raw scores, decide adoption. StepFun's move also signals China's open-weight models climbing global rankings and shifting from commercial export to ecosystem and standards export.
Source: StepFun / SegmentFault / Artificial Analysis
