Step3 is an open source multimodal macromodel developed by StepFun, whose core technical features include:
- Significantly optimized inference speed and performance using a hybrid model of experts (MoE) architecture with 32.1 billion total parameters (3.8 billion active parameters)
- Multimodal content generation capabilities supporting text, images and speech
- Model weights are available in both bf16 and block-fp8 storage formats to reduce hardware requirements
- Standardized API interface compatible with OpenAI and Anthropic
The project is licensed under the Apache 2.0 open source license, which allows developers to freely use, modify and distribute the code and model weights. This license gives developers the flexibility to use the code for both commercial and research purposes, fully reflecting StepFun's philosophy of promoting the development of artificial intelligence through open source.
This answer comes from the articleStep3: Efficient generation of open source big models for multimodal contentThe