Runway Opens the Weights of Its Video-Trained Robot Model, the Top World Action Model on Nvidia's Leaderboard but Fifth Overall
Runway released Praxis-1 on November 11 with open weights under Apache 2.0, including inference and fine-tuning code. The 27-billion-parameter model predicts video and robot actions together and scores 46.2% success on Nvidia's RoboLab-120 simulation benchmark, first among world action models and ahead of FLUX 3 Action (42.9%). On the leaderboard's overall table it ranks fifth, behind systems including an agentic language-model policy at 86.0%. Runway says it adapts to new robots with light fine-tuning.
Runway, best known for video generation, is giving away the robot model it built on that video. Praxis-1 is now available with open weights for commercial use, and on Nvidia's robotics leaderboard it is the strongest model of its type. The leaderboard also shows the limit of that claim: four systems of other kinds score higher.
What Runway released#
Runway made Praxis-1 available on November 11, after introducing it in September. Per Runway:
- It is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts future video and robot actions together, a "world action model."
- Weights, inference code and fine-tuning code, including LoRA, are released under Apache 2.0, on GitHub and Hugging Face with LeRobot support, and it runs on Nvidia hardware from data center GPUs to the Jetson Thor robot computer.
- Fine-tuned on Google's RT-1 robot, which it had not seen in training, it reached 80.0% success on the SimplerEnv test against 52.4% for the original RT-1 policy. Launch partners have run it on a humanoid, an industrial arm and a warehouse robot.
Where it ranks#
On Nvidia's RoboLab-120 simulation leaderboard, Praxis-1 succeeds in 555 of 1,200 episodes, or 46.2%, ranking first among world action models, ahead of Black Forest Labs' FLUX 3 Action at 42.9% and Nvidia's Cosmos3 Nano policy at 36.8%. In the overall table, which includes agentic language-model systems, it ranks fifth: the leader is listed at 86.0% and a vision-language-action model at 49.2%. The leaderboard flags several entries as using simulation data in training; Praxis-1 is not flagged. Nvidia's own GR00T N1.6 is listed at 7.2%; its newer GR00T 2, which Nvidia says ranks near the top of the field's leaderboards, does not appear on this table.
Limits#
- RoboLab-120 is a simulation benchmark. Nvidia says its rankings correlate strongly with a real-world benchmark, but simulation success is not deployment success.
- Runway's statement that Praxis-1 "outperforms every open-weight VLA" depends on which entries count as open-weight; the leaderboard's top entries are of different types and settings and are not directly comparable.
- The RT-1 and partner results are Runway's own.
- The leaderboard says submissions are temporarily paused; rankings were checked on November 12.