Windflow

Helios

PKU-YuanGroup · Released March 2026

Helios is a 14B long-video model that streams minute-scale video in real time while you change the prompt. Its authors reach real time without the usual shortcuts such as KV caches, sparse attention or quantization, and the weights are Apache-2.0.

See 3 places to try itGet the weightsCan I use it commercially?
Weights
Open weightsApache-2.0
Commercial use
Permissive license
Real-time
Real-time, per authorsNot independently measured
Where to try
3 placesHugging Face Space, Reactor, Windflow

01 Controls

What you control

  • A text prompt
  • Prompt changes while the video streams
  • An optional starting image

02 License

Can you use it commercially?

Permissive license

The weights' model card declares Apache-2.0, a permissive license that allows commercial use. Check what the model was built on, because a base model's terms can still apply.

Declared license: Apache-2.0 · Model card

A reading of the published license text, not legal advice.

03 Speed

Does it run in real time?

Real-time, per authors. Minute-scale video at 19.5 fps end to end.

Figures are the authors' own, with their settings and hardware. They are not comparable across models.

04 Setup

How to run it

  1. 01Clone PKU-YuanGroup/Helios, create a Python 3.11 conda environment, install PyTorch for your CUDA version, then run bash install.sh.
  2. 02Download the checkpoint you need from Hugging Face: Helios-Base, Helios-Mid or Helios-Distilled.
  3. 03Run the inference scripts in the repository, or load the model through Diffusers, which supports Helios.

Summarised from the project's own instructions. Follow the source for exact commands and versions.

05 Access

Where to try it

Listed alphabetically. Windflow maintains this directory and is one of the hosts; a listing does not promise capacity at any given moment.

06 Questions

Helios FAQ

Is Helios open source?

Helios publishes downloadable weights under Apache-2.0. The declared license is permissive.

Can I use Helios commercially?

The weights' model card declares Apache-2.0, a permissive license that allows commercial use. Check what the model was built on, because a base model's terms can still apply.

Does Helios run in real time?

Yes, according to its authors. Minute-scale video at 19.5 fps end to end. This directory has not measured it independently.

Where can I try Helios?

You can try it on Hugging Face Space, Reactor and Windflow.

What does Helios take as input?

You give it a text prompt, prompt changes while the video streams and an optional starting image.

How do I run Helios?

Following the official instructions: Clone PKU-YuanGroup/Helios, create a Python 3.11 conda environment, install PyTorch for your CUDA version, then run bash install.sh. Download the checkpoint you need from Hugging Face: Helios-Base, Helios-Mid or Helios-Distilled. Run the inference scripts in the repository, or load the model through Diffusers, which supports Helios.

What are alternatives to Helios?

Similar models in this directory include EdenLive (open weights, custom terms), Krea Realtime 14B (open weights, permissive license) and LongLive 2.0 (open weights, custom terms).

Similar world models

All 56 in the directory