Runway Open-Sources Praxis-1, a 27B Robot Model, Apache 2.0

Runway released open weights for Praxis-1, a 27-billion-parameter mixture-of-experts diffusion transformer for robot control, under Apache 2.0 with commercial use permitted. On Hugging Face, the company lists base, DROID and bimanual YAM variants alongside NVFP4-quantized builds whose routed experts run in 4-bit on SM120 GPUs.

The model pairs a 26.9B video backbone with a 16-layer action expert, totaling 27.2B parameters with roughly 2.8B active per token; one 80 GB GPU holds the 54.4 GB bf16 package. Each call predicts camera video and 24 action rows covering 1.6 seconds at 15 Hz.

On NVIDIA's RoboLab-120 leaderboard, Runway reports 46.2% success across 1,200 episodes, 3.3 points ahead of FLUX 3 Action at 42.9% and 18.2 points ahead of Physical Intelligence's π0.5 at 28.0%, with Nvidia's own Cosmos3 Nano at 36.8%. Fine-tuned on Google's RT-1 robot, which was not in Praxis-1's training set, the model reaches 80.0% on SimplerEnv against 52.4% for the original RT-1 policy, according to FourWeekMBA.

The training mix is 58% egocentric human video, 20% general video, and 18% real robot video, with fine-tuning on roughly 277 hours of successful DROID episodes for the Franka arm. "Web video matches teleoperated robot video as pretraining data for policy performance," Runway says.

The scores are the company's own, measured against other vendors' published figures rather than a third-party replication. The release follows Solaris in September and is Runway's first entry in our robotics coverage.