Open-Sora
HPC-AI TechOpen-Sora is HPC-AI Tech's open video-generation model family. The current release, Open-Sora 2.0, is an 11B-parameter model for text-to-video and image-to-video at up to 768px, trained for about USD 200K with the checkpoints and training code both published. Its own configs generate text-to-video directly, and an optional text-to-image-to-video pipeline uses FLUX.1 [dev] for the first frame.
Gap fill: the video models here are either closed or ship weights with at most inference or fine-tuning code; Open-Sora is the one video line that publishes its Apache-2.0 checkpoints together with the pretraining code. The training data itself is not released. Adoption is thin (under 4K Hub downloads a month across the family) despite 29.8K GitHub stars. The Open-Sora-v2 repository also redistributes, under their own licenses, the HunyuanVideo VAE every generation config loads and the FLUX.1 [dev] weights used only by the optional text-to-image-to-video pipeline. HPC-AI Tech's hosted Video Ocean product says it runs on "a superior model"; no newer Open-Sora release is named, so 2.0 is taken as current.
Openness
3 medium confidence- weights
- open(Open_Sora_v2.safetensors downloads from hpcai-tech/Open-Sora-v2 without a gate
- data
- described(the training data is described in the technical report
- code
- open(the pretraining scripts (scripts/diffusion/train.py) and docs/train.md are published
- license
- Apache-2.0(the Open-Sora 2.0 weights and the repository)+Tencent-Hunyuan-Community-License(the HunyuanVideo VAE that every text-to-video and image-to-video config loads, redistributed in the Open-Sora-v2 repository, and code the LICENSE bundles
- optional-pipeline
- FLUX.1-dev-Non-Commercial-License(FLUX.1 [dev] weights sit in the same Hugging Face repository and the README's download command fetches them, but only the optional text-to-image-to-video pipeline uses them
Open-Sora 2.0's own weights and training code are Apache 2.0, but every text-to-video and image-to-video run loads a HunyuanVideo VAE whose license excludes the EU, UK and South Korea, which caps it at use_bounded. FLUX.1 [dev] weights under a non-commercial license sit in the same repository for an optional text-to-image-to-video pipeline; the direct text-to-video path does not load them, so they do not govern. The training data is described, not released.
- https://huggingface.co/api/datasets/hpcai-tech/open-sora-pexels-45k recorded 2026-09-27
Dataset record for the Pexels sample: "disabled":true.
- https://huggingface.co/api/models/hpcai-tech/Open-Sora-v2?expand[]=downloads&expand[]=likes&expand[]=cardData&expand[]=gated&expand[]=lastModified&expand[]=createdAt recorded 2026-09-27
Hugging Face API record for hpcai-tech/Open-Sora-v2: cardData license apache-2.0, gated false, created 2025-03-04.
- https://huggingface.co/api/models/hpcai-tech/Open-Sora-v2/tree/main recorded 2026-09-27
File tree: Open_Sora_v2.safetensors, flux1-dev.safetensors, flux1-dev-ae.safetensors and hunyuan_vae.safetensors.
- https://huggingface.co/hpcai-tech/Open-Sora-v2/raw/main/README.md recorded 2026-09-27
Card: "You can also run direct text-to-video by:" the 256px and 768px configs, and separately "with the help of flux text-to-image model, we build a text-to-image-to-video pipeline".
- https://raw.githubusercontent.com/hpcaitech/Open-Sora/main/README.md recorded 2026-09-27
Model Download gives one command, "huggingface-cli download hpcai-tech/Open-Sora-v2 --local-dir ./ckpts", which fetches the whole repository, flux1-dev.safetensors included; Text-to-Video Generation says "To generate high quality videos, with the help of flux text-to-image model, we build a text-to-image-to-video pipeline" and lists direct text-to-video after it.
- https://ungh.cc/repos/hpcaitech/Open-Sora/files/main/configs/diffusion/inference/256px.py recorded 2026-09-27
Inference config: type="hunyuan_vae", from_pretrained="./ckpts/hunyuan_vae.safetensors", with T5 and CLIP text encoders; no FLUX weights are loaded.
- https://ungh.cc/repos/hpcaitech/Open-Sora/files/main/docs/train.md recorded 2026-09-27
docs/train.md: "we will use a 45k pexels dataset as an example."
- https://ungh.cc/repos/hpcaitech/Open-Sora/files/main/LICENSE recorded 2026-09-27
LICENSE body: "Apache License" ... "Copyright 2024 HPC-AI Technology Inc.", followed by bundled third-party terms including "THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA".
- https://ungh.cc/repos/hpcaitech/Open-Sora/files/main/README.md recorded 2026-09-27
README: "[2025.03.12] We released Open-Sora 2.0 (11B)" and "Fully open-source: checkpoints and training codes for training with only $200K."
Adoption
1 high confidenceAdoption is measured as Hugging Face downloads across the Open-Sora checkpoints, led by the 1.2 VAE and the 2.0 model. The project's GitHub following is far larger than its measured use.
- https://huggingface.co/api/models?author=hpcai-tech&search=Open-Sora&sort=downloads&direction=-1&limit=100 recorded 2026-09-27
Listing of hpcai-tech Open-Sora repositories sorted by downloads: OpenSora-VAE-v1.2 1321, Open-Sora-v2 1194, OpenSora-STDiT-v3 438, the rest under 320 each; 3816 in all.
Capability
1 medium confidenceOpen-Sora does not appear on the public video arena, where Mochi 1 places in the lower half. Its own report compares it with HunyuanVideo, but the map reads one public board per modality.
- https://artificialanalysis.ai/video/leaderboard/image-to-video recorded 2026-09-27
The image-to-video leaderboard lists no Open-Sora entry.
- https://artificialanalysis.ai/video/leaderboard/text-to-video recorded 2026-09-27
The text-to-video leaderboard lists no Open-Sora entry; "Mochi 1" (rank 73, Elo 1000) is on it.
Verified 2026-09-27