Praxis-1 is Runway's open-weight robot model: it predicts video and the movements a robot should make, both at once.
In plain words
Runway calls it a world action model: a 27-billion-parameter mixture-of-experts diffusion transformer that generates video and robot actions together. Because it learned how objects behave from huge amounts of video, it starts out ahead on a robot body it has not met, before any fine-tuning. Runway introduced it in September and has now released the weights under the Apache 2.0 licence, with code for running and fine-tuning it, LoRA included. The base checkpoint lacks robot-specific action parts, so you add those for your machine, then fine-tune.
A quick example
Imagine a lab with a two-armed robot that Praxis-1 has never seen. The team downloads the base weights, adds the parts that match their arms and cameras, then fine-tunes with LoRA on recordings of their own tasks instead of retraining the whole model.
Where you'll meet it
- Robot control research and products.
- Adapting one model to different robot bodies.
Related
Sources
- DocsPraxis-1 is now available, Runway · read 12 Nov 2026
- Docsrunwayml/praxis-1-base, Runway (Hugging Face) · read 12 Nov 2026
Checked by WhatIsAI against the sources above · 12 Nov 2026. Spotted an error? Report it.