A model developer with several modalities
StepFun develops the Step model family and provides an API platform for developers. Its work spans language, visual understanding and speech. Those capabilities belong to different models and access channels, rather than to a single interchangeable product. [1]
Model design meets inference systems
The Step3 public-weight release combined a mixture-of-experts language model with visual inputs. StepFun reported 321 billion total parameters and 38 billion active parameters, and described co-designing attention and distributed inference to reduce decoding costs. Those efficiency comparisons are the developer’s measurements. [2]
Public artifacts and their limits
Step3’s official repository identifies both code and weights as Apache 2.0 licensed and provides deployment guidance. The July 31 release record is distinct from the earlier conference unveiling and technical-paper submission. Its research account also reports weaknesses in some coding tasks and a tradeoff between textual reasoning and visual perception. [3][2]
Sources
- StepFun
StepFun · retrieved Oct 23, 2026
- Step3: Cost-Effective Multimodal Intelligence
StepFun · Jul 31, 2025
- stepfun-ai/Step3
StepFun · retrieved Oct 23, 2026