On October 15, StepFun (阶跃星辰) officially open-sourced the complete weights of its new flagship base model Step 5 Preview. First announced on September 20 and aimed at AI coding, software engineering, knowledge work and finance, the model boosts performance on real-world Agentic tasks and has now fully opened third-party interfaces.
Step 5 Preview uses a sparse Mixture-of-Experts architecture with 600 billion total parameters and only 27 billion activated per token, supports a 1-million-token context window and natively handles both text and vision input. On the Artificial Analysis Intelligence Index it scores 44, ranking among the global top three open-source models, behind Qwen3.8 Max and GLM-5.3.
StepFun estimates that for the same intelligent level, a single task costs about 0.71 US dollars (roughly 4.76 yuan) — about one-eighth of Claude Opus 5's 5.86 dollars. The team argued that the next stage of scaling laws is not just more compute but more efficient compute. China's open-source model downloads have surpassed 10 billion times, and the logic of open-weight models is to build ecosystem advantages through global developers.
The open-sourcing lowers the barrier for small and medium enterprises and individual developers, as a model ranking among the global top three with one-eighth the cost of a top model makes AI development markedly more accessible.
Source: CNMO / NetEase / Toutiao
