Open-Sora
Apache-2.0 11B text-to-video and image-to-video model from HPC-AI Tech with training and inference code.
Open-Sora 2.0 generates 256px and 768px video from text or an image.
Open-Sora is an open effort to reproduce a Sora-class video model, now at version 2.0 with 11 billion parameters. It suits research teams and well-equipped developers who want to run and study a large open video model.
What it is
Open-Sora 2.0 handles text-to-video and image-to-video in one model at 256px and 768px, in several aspect ratios. The repo includes inference scripts, a Gradio demo and weights on Hugging Face and ModelScope.
Who it's for
It suits teams with H100-class GPUs and Python skills. It does not suit laptops or single consumer cards, since the README reports 52 to 60GB of peak memory for one GPU.
Key features
T2V and I2V in one
A single 11B model covers text-to-video and image-to-video.
256px and 768px
Two resolutions with 16:9, 9:16, 1:1 and 2.39:1 aspect ratios.
Frame control
Frame counts follow a 4k+1 rule and stay under 129.
Gradio demo
A Gradio folder and a Hugging Face Space demo the model.
Open weights
Checkpoints are on Hugging Face and ModelScope for download.
Multi-GPU runs
The README shows multi-GPU runs that cut time per clip.
Up and running in 3 steps
Our honest verdict
What's great
- Apache-2.0 repository with the full 11B model and code in one place.
- The README states human-preference results on par with HunyuanVideo 11B and Step-Video 30B.
- Multi-GPU inference is documented, with per-GPU memory falling to about 44 to 48GB.
- Weights come from two mirrors, so downloads do not depend on one host.
Where it falls short
- Peak memory is 52 to 60GB on one GPU, so you need data-centre hardware or several cards.
- Setup involves a conda environment, xformers and a flash-attn build, which must be built against your CUDA version.
- The latest model news in the README dates from March 2025, and clips are short at 256px or 768px.
- The README gives no clip length in seconds and no separate weights licence note.
Pricing & license
Free under Apache-2.0. The cost is H100-class GPU time or hardware; the repo lists no paid hosted plan.
Open remote AI-training roles
870 open remote AI-training roles, updated November 14, 2026; most pay $20–$170 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
These reviewed AI tools do the same job instantly — from our scored directory.
Use it, or have us set it up
FAQ
What GPU does Open-Sora 2.0 need?
The README reports 52.5GB peak memory at 256x256 and 60.3GB at 768x768 on one GPU, measured on H100 and H800 cards. Multi-GPU runs use about 44 to 48GB per card.
Is Open-Sora free for commercial use?
The repository lists Apache-2.0 and the README adds no separate weights licence. Confirm the model card on Hugging Face before shipping a product.
How is Open-Sora different from a hosted tool like Runway?
You download and run the model on your own hardware, with no credits or queue, but you handle installation, GPU capacity and prompt tuning yourself.
