Runway Opens Praxis-1 Under Apache 2.0, and NVIDIA's Own Leaderboard Puts It Fifth

Runway released the weights of Praxis-1, a robot-control model it calls a world action model, under the Apache 2.0 licence on Nov. 11. On the RoboLab-120 simulation leaderboard that NVIDIA hosts, the model sits fifth at 46.2%, behind four entries that include a policy built on a language-model agent at 86.0%.

Runway says the licence covers commercial use and that the release includes weights plus inference and fine-tuning code, according to its launch post. The code is on GitHub and the weights are on Hugging Face with LeRobot support.

What is in the Praxis-1 release

The model is a mixture-of-experts diffusion transformer that predicts video and actions together. Runway's research page puts the video backbone at 26.9B parameters, about 2.8B active per token, plus a 314M action expert, for 27.2B in total.

The bf16 weights are 54.4 GB and run on one 80 GB GPU. With NVFP4-packed experts they shrink to 17.5 GB. A call takes about 0.52 to 0.55 seconds on one H100 by default and about 0.12 to 0.13 seconds with an opt-in real-time chunking mode.

Andy Chen, head of robotics at Runway, is quoted by The Robot Report as saying: "A strong base model goes a long way." It(https://www.therobotreport.com/runways-praxis-1-world-action-model-for-robotics-is-now-available/) lists launch partners Noble Machines, Standard Bots and Ultra.

Where RoboLab-120 puts it

RoboLab-120 is a simulation benchmark of 120 tasks with 10 trials each. Praxis-1 solved 555 of 1,200 episodes. Runway calls it the top-performing world action model on the board, and the NVIDIA leaderboard is consistent with that: FLUX 3 Action scores 42.9% and Cosmos3-Nano-Policy 36.8%.

1. AMP1 (agentic LM)86 %
2. BiMind v0.5 (VLA)49.2 %
3. Praxis-1 (WAM)46.2 %
4. FLUX 3 Action (WAM)42.9 %
5. Cosmos3-Nano-Policy (WAM)36.8 %
6. pi0.5 (VLA)28 %

Source: NVIDIA Research RoboLab-120 leaderboard, accessed 2026-11-12

Runway also says Praxis-1 outperforms every open-weight VLA, a vision-language-action model. The leaderboard lists BiMind v0.5, a VLA, at 49.2%, above Praxis-1. The page does not say whether BiMind's weights are public, so the claim cannot be checked there. The page also says submissions are paused, and it carries no update date.

What the real-robot numbers show

Test	Praxis-1	Comparison
Franka FR3, 100 real episodes	88	80 (FLUX 3 Action)
RT-1 robot, SimplerEnv simulation	80.0%	52.4% (original RT-1 policy)

Both rows are Runway's own figures from its research page, and the RT-1 test is simulated. The authors list their own limits: reaching for look-alike objects is the most common real-robot failure, vague instructions are harder, and chaining several steps from one prompt remains hard.

The training mix is 58% egocentric human video, 20% general video and 18% real-robot video by token share. Runway's research page does not report a head-to-head of teleoperated data against web video, so the launch claim that web video matches it is not backed by a table there.

Earlier on this site, Berkeley and NVIDIA reported a coding-agent approach to robot tasks, the same family as the agentic entries at the top of this board. Deployment numbers elsewhere in robotics have been thin too: Agility's $2.5 billion listing rested on $1.8 million of revenue. Runway has not published hosted-inference pricing. Runway's launch post says Praxis-1 can be tried on Runway Dev now.