Market · AI in practice
Runway opens the weights of Praxis-1, a model that predicts video and action together
Runway released Praxis-1 with open weights under an Apache 2.0 license, including for commercial use. It is a 27 billion parameter mixture of experts diffusion transformer that predicts video and actions together. Partners including Noble Machines, Standard Bots and Ultra were already running it on their own hardware.
What the sources report
Praxis-1, which Runway had already presented, is now out with open weights under an Apache 2.0 license, including for commercial use. It is a 27 billion parameter mixture of experts diffusion transformer that predicts video and actions together. Launch partners including Noble Machines, Standard Bots and Ultra were already running it on their own hardware. The same model is on GitHub, Hugging Face and vLLM-Omni.
On the RoboLab-120 leaderboard, according to Runway, Praxis-1 reached 46.2% success across 1,200 episodes, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and π0.5 at 28.0%. Fine-tuned on Google's RT-1, which was outside its training data, it reached 80.0% on SimplerEnv, compared with 52.4% for the original policy, also a figure from Runway.
My take
Open weights matter here because the team that has the machine can run and fine-tune the model on it.