Runway has released Praxis-1, a new AI model designed to translate knowledge from video pre-training into real-world robotic control. The model is available under the Apache 2.0 license, allowing developers to run and fine-tune it on their own hardware.
The architecture relies on a 270-billion parameter video foundation model. Runway provides an NVFP4 quantized version optimized for the Nvidia RTX 5090 GPU. The source code and model weights are accessible via GitHub and Hugging Face.
In testing on the RoboLab-120 benchmark, Praxis-1 completed 555 out of 1200 scenarios, achieving a 46.2% success rate. The model predicts both future scene states and the specific actions required to achieve them.
Compared to systems powered by GPT-6 Astra, Praxis-1 operates 15.6 times faster on identical workstations. Additionally, the operational cost per task is reduced to 1/51 of the cost of the alternative system.
