Recreating Sora's Video Generation, Trained Entirely on Ascend Hardware
You've probably seen the Sora demo videos—minute-long, photorealistic clips generated from a single text prompt. The catch? OpenAI hasn't open-sourced it, and the hardware requirements for training something like it are steep. If you've ever wanted to dig into how text-to-video generation actually works under the hood, you're stuck reverse-engineering from papers and hoping the community figures it out. That's exactly the gap Open-Sora Plan is trying to fill.
Open-Sora Plan is an open-source project from PKU-YuanGroup (with contributions from Huawei, Pengcheng Lab, and community partners) that aims to reproduce Sora's capabilities in a simple, scalable codebase. The current V1.5 release has a notable claim attached to it: it was trained entirely on Huawei Ascend hardware. No NVIDIA GPUs required. That's a meaningful statement for anyone interested in hardware diversity in AI research.
What It Does
At its core, Open-Sora Plan is a text-to-video generation system. You give it a prompt, it generates video. But the project's real focus is on being a reproducible, scalable research codebase rather than just a polished demo. The README describes it as "a simple and scalable repo" designed to reproduce Sora—which the project cheekily refers to as "ClosedAI" given the lack of open weights or training details.
The project is a joint effort between academia and industry, led by the Peking University-TuZhan AIGC Joint Laboratory. The V1.5 version is described as "Ascend pure-blood" (昇腾纯血版), meaning the entire training pipeline runs on Huawei's Ascend AI accelerators. This is a significant technical achievement because most large-scale video generation models are trained on NVIDIA hardware with CUDA, and porting that training stack to a different accelerator architecture is non-trivial.
The repo includes links to three arXiv papers covering different components: the main Open-Sora Plan architecture, a system called Helios for long video generation, and WF-VAE for video compression. So this isn't just a single model—it's a collection of research contributions around video synthesis.
Why It's Cool
Let's be honest about what makes this project stand out:
-
Hardware independence is a big deal. Most open-source AI projects assume you have NVIDIA GPUs. Open-Sora Plan V1.5 being trained entirely on Ascend means there's now a reference implementation for large-scale video generation that doesn't depend on CUDA. If you're working in regions where NVIDIA hardware is restricted, or you're just interested in alternatives, this matters.
-
It's genuinely reproducible. The project's stated goal is reproducibility, not just impressive demos. That means the codebase is designed to be straightforward enough that researchers can actually run it and build on it, which is more than you can say for many flashy AI releases.
-
The community angle. This is a collaboration between a university lab, a company (TuZhan), a national lab (Pengcheng), and open-source contributors. That's a healthy mix of research rigor and practical engineering. The project explicitly invites pull requests and new collaborators.
-
It's actively developed. The README shows a busy repository—lots of contributors, merged PRs, and recent activity. The project also mentions a paper released in March 2026 about Helios, a model that achieves 19.5 FPS video generation on a single H100 without conventional anti-drifting techniques. That's a concrete performance claim you can verify.
-
The papers are there. If you want to understand how it works rather than just run it, the arXiv links give you the full technical details on the architecture, the video compression approach, and the long-video generation method.
How to Try It
The project is on GitHub at pku-yuangroup/open-sora-plan. The README doesn't include quick-start installation commands in the portion we have, but here's what you should know:
-
Check the repository's README directly for the most current setup instructions, dependencies, and model weights. The project is moving fast, so the README is your best source of truth.
-
Look at the arXiv papers linked from the repo to understand the architecture before you run anything. The WF-VAE paper (arXiv:2411.17459) covers the video compression component, and the Helios paper (arXiv:2603.04379) describes the long-video generation system.
-
Join the community channels. The README links to a Discord server and a WeChat group for discussion. If you hit issues getting things running, those are the places where maintainers and other users hang out.
-
Consider contributing. The project explicitly welcomes pull requests and is looking for algorithm engineers. If you're working on video generation or just want to help with an Ascend-based training stack, this is a project that wants your involvement.
The repo also has badges for model experience platforms, so there may be hosted demos available—check the Modelers link in the README if you want to see outputs before committing to a local setup.
Final Thoughts
Open-Sora Plan is best for researchers and engineers who want to work with a serious, open text-to-video codebase—especially if you're interested in training on non-NVIDIA hardware. It's not a turnkey product; it's a research repository with real complexity and a steep learning curve. But that's also its strength. The project is tackling one of the hardest problems in generative AI, doing it in the open, and doing it on hardware that most of the field has ignored. If you've been waiting for a reason to explore video generation outside the CUDA ecosystem, this is a solid place to start.
Follow @githubprojects for more developer tools and open source projects.