Step3's model weights are stored in bf16 and block-fp8 formats, which optimize the storage and computation efficiency of model weights. Developers can download these weights from the Hugging Face platform and select the appropriate format according to their needs. This optimization reduces hardware requirements and is particularly suitable for deployment in multi-GPU environments, such as using four A800/H800 GPUs with 80GB of video memory for optimal performance.
This answer comes from the articleStep3: Efficient generation of open source big models for multimodal contentThe
May not be reproduced without permission:AI productivity tools " Step3's model weights are stored in bf16 and block-fp8 formats
English 
简体中文 
日本語 
Deutsch 
Português do Brasil