# Krea Realtime 14B

A 14-billion-parameter text-to-video model that Krea distilled from Wan 2.1 14B with Self-Forcing, so it generates video frame by frame while you change the prompt. It can also restyle a live video, webcam or canvas input. The weights are Apache-2.0, but the official inference code is CC BY-NC-SA 4.0.

Organization: Krea. Released: October 2025. Kind: Live video.

## What you control

- A text prompt, which you can change while it generates
- A video, webcam stream or canvas drawing to restyle

## License

Weights: Available (https://huggingface.co/krea/krea-realtime-video). Declared license: Apache-2.0.
The weights' model card declares Apache-2.0, a permissive license that allows commercial use. The official inference code in krea-ai/realtime-video is under CC BY-NC-SA 4.0, so a commercial product needs its own inference code; the Wan 2.1 base model is Apache-2.0.

## Real-time

Real-time, per authors. 11 fps text-to-video with 4 inference steps, and first frames within about a second, per the README and model card.

## How to run it

1. Install the inference code with uv sync, then an attention backend (FlashAttention, or the SageAttention wheel bundled in the repository) and FFmpeg.
2. Download the Wan2.1-T2V-1.3B base files and krea-realtime-video-14b.safetensors from Hugging Face.
3. Start release_server.py to stream frames live over a WebSocket, or run sample.py for offline clips.

## Where to try

No hosted destination verified.

## Sources

- [Inference code](https://github.com/krea-ai/realtime-video)
- [Weights and model card](https://huggingface.co/krea/krea-realtime-video)
- [Technical blog post](https://www.krea.ai/blog/krea-realtime-14b)

## Questions

### Is Krea Realtime 14B open source?

Krea Realtime 14B publishes downloadable weights under Apache-2.0. The declared license is permissive.

### Can I use Krea Realtime 14B commercially?

The weights' model card declares Apache-2.0, a permissive license that allows commercial use. The official inference code in krea-ai/realtime-video is under CC BY-NC-SA 4.0, so a commercial product needs its own inference code; the Wan 2.1 base model is Apache-2.0.

### Does Krea Realtime 14B run in real time?

Yes, according to its authors. 11 fps text-to-video with 4 inference steps, and first frames within about a second, per the README and model card. This directory has not measured it independently.

### Where can I try Krea Realtime 14B?

No hosted version was verified; run the published weights yourself.

### What does Krea Realtime 14B take as input?

You give it a text prompt, which you can change while it generates; and a video, webcam stream or canvas drawing to restyle.

### How do I run Krea Realtime 14B?

Following the official instructions: Install the inference code with uv sync, then an attention backend (FlashAttention, or the SageAttention wheel bundled in the repository) and FFmpeg. Download the Wan2.1-T2V-1.3B base files and krea-realtime-video-14b.safetensors from Hugging Face. Start release_server.py to stream frames live over a WebSocket, or run sample.py for offline clips.

### What are alternatives to Krea Realtime 14B?

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

Checked 2026-11-03. Corrections: https://github.com/windflow-labs/world-model-directory