Runway Opens Praxis-1, a 27B Video-Trained Robot Brain

Robots have long been held back by a data problem: teaching a machine to handle a mug or a cutting board usually means collecting thousands of hours of costly, hand-guided demonstrations. Runway, better known for video generation, is betting that ordinary video can do much of that teaching. On November 11, 2026, the company released the weights of its Praxis-1 robotics model to everyone.

What Runway released

According to Runway’s announcement, Praxis-1 is a world action model, meaning it predicts video and robot actions together. It is a 27-billion-parameter mixture-of-experts diffusion transformer, built on the same large-scale video pretraining behind Runway’s general world models. Runway first introduced it in September, with launch partners Noble Machines, Standard Bots and Ultra testing it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot.

The new step is openness. The weights, inference code and fine-tuning code are available under the Apache 2.0 license, which Runway says includes commercial use. The model is on GitHub, on Hugging Face with native LeRobot support, and in vLLM-Omni, with hosted inference on Reactor.

How it scores

Runway reports that Praxis-1 is the top-performing world action model on NVIDIA’s RoboLab-120 leaderboard and outperforms every open-weight vision-language-action model. Across 1,200 episodes it reached 46.2% success. The company lists FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and pi0.5 at 28.0%. These are Runway’s own figures as of November 2026, and a 46% success rate is a reminder that general-purpose robot manipulation is still far from solved.

A second result speaks to flexibility. Fine-tuned on Google’s RT-1 robot, which never appeared in Praxis-1’s training data, the model reached 80.0% success on SimplerEnv, compared with 52.4% for the original RT-1 policy.

Why video helps robots

The core idea is that video already shows how objects behave and how tasks unfold. Runway says its research found that web video matches teleoperated robot video as pretraining data for policy performance. If that holds up more widely, the bottleneck shifts from collecting robot-specific demonstrations to how much general video a model can learn from, which is a far larger and cheaper pool.

That head start matters when a team has a new robot. Runway says a single Praxis-1 checkpoint adapted to a variety of embodiments with light fine-tuning, and that LoRA lets developers adapt it to their own cameras, tasks and body without retraining the full model. It runs across NVIDIA hardware, from H100 and B200 data center GPUs to RTX 5090 and RTX PRO 6000 workstation cards, and on Jetson Thor on the robot itself. NVIDIA’s NeMo Automodel project has hands-on tutorials for post-training it.

Why open weights

Runway argues that physical AI needs both hardware and software investment, and that hardware developers benefit from being able to run, inspect and adapt a model on their own terms. The release builds on its role as a founding member of the Cosmos Coalition, which works on open world models, and on its collaboration with NVIDIA on shared foundations for developers.

What it means

For small robotics teams and university labs, a commercially usable 27B model removes a major barrier: they no longer need to train a foundation policy from scratch or negotiate access to a closed one. For larger manufacturers, it adds a baseline to compare against, and open weights make independent testing easier than with closed systems.

Caveats are worth keeping in mind. The leaderboard and transfer numbers come from the developer and benchmarks rather than from long deployments. A 27B model also needs serious compute, so the on-robot story depends on hardware such as Jetson Thor. Real factory or warehouse reliability will be shown by partners’ deployments, not by benchmark tables.

Conclusion

Praxis-1 is a notable test of whether video-first training can make robot brains cheaper to build and easier to share. By opening the weights, Runway invites the wider robotics community to find out.

Sources