Step-3

model paper
StepFun
321B parameter MoE model (38B active) designed for high-efficiency multimodal reasoning with cost-effective decoding.
AnnouncementHuggingFacePaper (arXiv)GitHub

Outputs 2

Step-3

model

321B parameter MoE model (38B active) designed for high-efficiency multimodal reasoning.

AnnouncementHuggingFace
Architecture MOE
Parameters 321B
Active params 38B

Released Jul 31, 2025.

Step-3: Large yet Affordable Model-System Co-design

paper

Flagship LLM paper detailing cost-effective decoding for the Step-3 MoE model.

Paper (arXiv)GitHub

arXiv: 2507.19427

moereasoningmultimodalefficiencyscaling