Open-Sora Plan

PKU-YuanGroup's leading open-source Sora reproduction (12.2K★; Li Yuan, Shenzhen Graduate School, with the PKU–Rabbitpre joint lab): v1.5.0 (June 2025) pretrains an 8B-scale sparse "SUV" video DiT from scratch on 40M video samples over the group's WF-VAE (CVPR 2025), matching open-source HunyuanVideo — trained and inferred entirely on Huawei Ascend 910-series NPUs. Successors: OSP-Next (May 2026; sparse sequence parallelism, HiF8 quantization, RL post-training) and the Helios real-time long-video generator (~2K★ in 3 months). Sibling ConsisID identity-preserving T2V (CVPR 2025 Highlight) ships in HF diffusers.