Open-Sora 2.0

Open-Sora 2.0 was released on March 12, 2025 by the hpcaitech/Open-Sora team (originating from Peking University's effort to reproduce Sora-style video generation with RabbitPre). It is an 11B-parameter model for text-to-video and image-to-video generation at up to 768x768 resolution and 5-second clips, and was reported on par with HunyuanVideo and the 30B Step-Video on VBench and human preference evaluations.

The project's headline claim is cost efficiency: a commercial-level video generation model trained for only about $200K, described as 5-10x cheaper than comparable models, achieved through joint optimization of data curation, training strategy, and infrastructure. Fully open-source, it ships checkpoints, training code, data-preparation pipelines, and technical reports, and continues the Open-Sora line that began with the 1.0 release in March 2024. Check the repo for the exact license before commercial use.