Runway, the company that made its name generating video, has published the complete weights of its robotics model.
On November 11, Runway announced that Praxis-1 is now open source under the Apache 2.0 license, with no restrictions on commercial use, fine-tuning or redistribution. Weights, inference code, full fine-tuning code and LoRA fine-tuning code are all available on GitHub and Hugging Face.
Imagine First, Then Move
Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer. Runway calls it a world-action model: before acting, the model first generates roughly the next 1.6 seconds of what it expects to see, then derives the robot's movements from that imagined future.
The design targets an old problem in robotics. Traditional control policies map sensor readings directly to motor commands with no anticipation of the physical world, while pure video models can imagine but not act. Folding both into one model lets a robot think through an outcome before committing its hands.
Training runs in four stages: general video pretraining, mid-training on embodied video dominated by first-person human footage, joint video-and-action training, and finally fine-tuning for a specific robot. According to the release materials, each robot only adds thin input and output layers while the backbone and action expert stay shared, which explains the low adaptation cost across machines.
The Score, and Its Boundary
On Nvidia's RoboLab-120 simulation benchmark, the DROID-fine-tuned version of Praxis-1 solved 555 of 1,200 episodes, a 46.2 percent success rate that ranks first among world-action models, ahead of FLUX 3 Action at 42.9 percent, Cosmos3 Nano at 36.8 percent and pi-zero-point-five at 28.0 percent.
Transfer results add context: fine-tuned onto Google's RT-1 robot, a platform it had never seen in training, Praxis-1 reached 80.0 percent on SimplerEnv against the original policy's 52.4 percent.
The leaderboard also frames the claim. Independent coverage noted that on the combined table including all system types, Praxis-1 ranks fifth, with the top entry, an agentic language-model system, listed at 86.0 percent. Simulation success is not deployment success, a caveat Nvidia itself prints on the leaderboard page.
From the Server Room Onto the Robot
Deployment requirements stay modest. The full bfloat16 model is 54.4 GB and runs on a single 80 GB GPU; a 4-bit quantized version shrinks to 17.5 GB, small enough for an RTX 5090 workstation card or the Jetson Thor computer mounted on the robot itself. Runway's stated inference latency is roughly 0.55 to 0.6 seconds per step.
Three hardware partners were already running the model at release: a legged humanoid, a six-axis industrial arm and a dual-arm warehouse station. Runway's showcase task is wooden-plank packing, where half an hour of demonstrations produced 15 successful completions in 15 attempts. That figure is the company's own and has not been independently replicated.
Runway promised the open release in September and delivered within two months. Video pretraining for robot control now has a complete, downloadable implementation.
For small teams, the meaning is access: no need to train from scratch, just fine-tune with your own robot's data. For the industry, the competitive coordinates have shifted. Companies that generate images and video are extending generation into the physical world, and with one player's cards face up, rivals must either open their own weights or publish stronger closed results.
Disclaimer: This article is based on public information and is for reference only. It does not constitute any investment advice.
