StepFun open-sourced Step-Video-T2V—a 30B-parameter DiT trained with a 16x16x8 Video-VAE and Video-DPO to generate 204-frame clips—alongside Step-Audio, a 130B open-weight speech and voice interaction framework.
StepFun (Step-Video & Step-Audio)
Also known as: StepFun · Step-Video · Step-Video-T2V · Step-Audio · Step1X-Edit · Step-3
Plain English
The Shanghai AI lab founded by former Microsoft Research Asia Chief Scientist Jiang Daxin—known for open-sourcing the 30B Step-Video-T2V generator, the 130B Step-Audio voice model, Step1X-Edit image editor, and the hardware-co-designed Step-3 multimodal MoE.
Technical Overview
StepFun develops large open-weight multimodal foundation models across video, audio, image editing, and reasoning: Step-Video-T2V (30B DiT with a 16x16x8 deep Video-VAE and Video-DPO), Step-Audio (130B dual-codebook speech-to-speech model), Step1X-Edit (VLM-conditioned DiT image editor), and Step-3 (321B/A38B MoE using Multi-Matrix Factorization Attention and Attention-FFN Disaggregation).
Timeline (4)
2025
Release- Paper
StepFun released Step-Video-TI2V (30B), extending its open-weight video foundation backbone to text-conditioned image-to-video generation with explicit motion-magnitude scoring and camera-stability control.
- Release
StepFun open-sourced Step1X-Edit, coupling a Qwen-VL multimodal encoder with a DiT image denoiser to perform natural-language reference image editing locally.
- Paper
StepFun published Step-3 (321B total / 38B active MoE VLM), introducing Multi-Matrix Factorization Attention (MFA) to cut KV cache traffic and Attention-FFN Disaggregation (AFD) to pipeline attention and expert layers across separate GPU pools.
See also (10)
Open full graphStepFun (Step-Video & Step-Audio)