Runway has released Praxis-1, its 27-billion-parameter world action model for robotics, as open weights under the Apache 2.0 licence — weights, inference code and fine-tuning code included, for commercial use. Why it matters: robot builders can now download, inspect and adapt a frontier robot policy on their own hardware instead of waiting for API access.
November 11, 2026
What happened
Runway introduced Praxis-1 in September as its first world action model, tested until now only with selected partners. On November 11 the company made the same model available to everyone: the weights are on Hugging Face with native LeRobot support, the code is on GitHub, it runs in vLLM-Omni, and a hosted version is available on Reactor for teams without their own GPUs.
Launch partners Noble Machines, Standard Bots and Ultra have already run it on their own hardware — a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot, respectively.
How Praxis-1 works
Praxis-1 is a mixture-of-experts diffusion transformer built on the same large-scale video pretraining behind Runway’s general world models. A video backbone of about 26.9 billion parameters (roughly 2.8 billion active per token) imagines what the scene will look like next, while a much smaller action expert — about 314 million parameters — reads that imagined future and outputs the motion to match. Each call returns 1.6 seconds of actions: 24 action rows at 15 Hz. The full model is 54.4 GB in bf16 and runs on a single 80 GB GPU, from data-centre H100 and B200 cards to an RTX 5090 workstation or a Jetson Thor computer on the robot itself, Runway says.
What the numbers do — and do not — show
On NVIDIA’s RoboLab-120 benchmark, Runway reports Praxis-1 at 46.2% success across 1,200 simulated episodes, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and π0.5 at 28.0%. Fine-tuned onto Google’s RT-1 robot — an embodiment that never appeared in its training data — it reached 80.0% on SimplerEnv, against 52.4% for the original RT-1 policy, and teams can adapt it to new hardware with full fine-tuning or the cheaper LoRA method.
Treat those figures as vendor-reported: they are Runway’s own results, dated November 2026, with no independent testing yet. Runway’s own research also flags real limits — the model can reach for the wrong object when two items look alike, struggles with vague instructions and long multi-step tasks, and still needs data-centre-class hardware for the full model.
Why it matters
Robotics has had strong closed models and weaker open ones. An Apache 2.0 release at the top of a public leaderboard changes the starting point for every small robotics team: instead of training a policy from scarce teleoperation data, they can start from a model that already learned how objects behave from video, then fine-tune it on their own robot. For subscribers, the practical test is the next few months — whether independent teams reproduce the RoboLab results on real hardware, and how the model handles the messy, look-alike, badly-lit tasks that simulators leave out. Everything AI, tested means watching the reproductions, not just the launch numbers.
Frequently asked questions
Is Praxis-1 free to use commercially?
Yes. Runway released the weights and the inference and fine-tuning code under Apache 2.0, which allows commercial use.
Where can I download Praxis-1?
The weights are on Hugging Face with LeRobot support, the code is on GitHub, and it also runs in vLLM-Omni or as hosted inference on Reactor.
What hardware does Praxis-1 need?
The full model is 54.4 GB in bf16 and is designed for a single 80 GB GPU. Runway lists H100 and B200 data-centre cards, RTX 5090 and RTX PRO 6000 workstations, and Jetson Thor on-robot computers.
How good is Praxis-1 compared with other robot models?
In Runway’s own RoboLab-120 testing it scored 46.2%, ahead of FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%) and π0.5 (28.0%). Those are company-reported figures with no independent verification yet.
Sources: Runway (Praxis-1 availability announcement and Praxis-1 research report), SXF Brief, MadRobot.