Runway made Praxis-1 available with open weights and code, saying the 27B robot model leads its categories on RoboLab-120.

Runway opens Praxis-1 robotics model under Apache 2.0

Runway opens Praxis-1 robotics model under Apache 2.0

Image: AI-generated editorial illustration

Runway announced Praxis-1 on November 11, making its robotics model available with open weights under Apache 2.0. The company says the 27-billion-parameter system turns video pretraining into robot control and, in its own evaluation, solved 46.2% of RoboLab-120 episodes.

Runway announced Praxis-1 on November 11, releasing its robotics model with open weights under the Apache 2.0 license and positioning it as a world-action model for robot control. The company says the system turns video pretraining into control for real robots, and that it is available to everyone through GitHub, Hugging Face, vLLM-Omni, and hosted inference on Reactor.

The release gives founders and robotics builders a clearer look at how Runway is packaging its video foundation work for physical systems. Runway says Praxis-1 uses a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together, with an architecture split into a video backbone, an action expert and thin per-robot projections. The company’s research post says each call takes an instruction, camera frames and the robot’s current pose, then returns predicted future video plus 1.6 seconds of actions at 15 Hz.

What is Runway shipping with Praxis-1?

Runway is shipping Praxis-1 as an open-weight robotics model for robot control, alongside inference and fine-tuning code. The company says the release includes support for full fine-tuning and LoRA, and that the weights and code are available under Apache 2.0, including commercial use.

In its launch materials, Runway also says the model runs across NVIDIA hardware ranging from H100 and B200 data-center GPUs to RTX 5090, RTX PRO 6000 and Jetson Thor. The company framed the open-weight release as a way to give hardware developers flexibility to run, inspect and adapt a world action model on their own terms.

Runway said it first introduced Praxis-1 in September as its first world action model, built on the same large-scale video pretraining behind its general world models. It also said launch partners Noble Machines, Standard Bots and Ultra had already been running Praxis-1 on their own hardware, including a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot.

How did Praxis-1 perform on RoboLab-120?

Runway says Praxis-1 solved 555 of 1,200 RoboLab-120 simulation episodes, a 46.2% success rate, under the benchmark’s official protocol. In the company’s comparison table, that result is above FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and π0.5 at 28.0%.

Runway’s launch post and research page say the model ranks first among world action models and among open-source robot foundation models on NVIDIA’s RoboLab-120 leaderboard. That claim is narrower than an overall first-place finish: NVIDIA’s leaderboard lists Praxis-1 fifth overall, with 555 of 1,200 episodes solved and 27.3 billion parameters.

Runway also said the model was evaluated on a real Franka FR3 robotic arm operated by Positronic, in addition to simulation testing. In a separate fine-tuning result, Runway said adapting Praxis-1 on Google’s RT-1 robot, which it said did not appear in training data, reached 80.0% success on SimplerEnv versus 52.4% for the original RT-1 policy.

Why does Runway say open weights matter for robotics?

Runway’s stated rationale is that robotics developers need the ability to run, inspect and adapt control models on their own systems. That is why the company is making Praxis-1 available with open weights rather than only through an API.

The release also ties into Runway’s role in the Cosmos Coalition with NVIDIA. Runway says the model is part of that work and that it is designed to run on a broad spread of NVIDIA hardware, from data-center GPUs to edge devices like Jetson Thor.

The company says its technical report uses web video paired with teleoperated robot video as pretraining data for policy performance. It also details internal scaling: the video backbone has 26.9 billion parameters with about 2.8 billion active per token, while the action expert has about 314 million parameters. Runway says the full model totals 27.2 billion parameters, occupies 54.4 GB in bf16 and runs on one 80 GB GPU.

What should builders watch next?

For robotics teams, the immediate implication is that Praxis-1 is not just a benchmark result but a downloadable model with code, commercial licensing and fine-tuning support. That makes it easier for developers to test whether a video-first policy can be adapted to their own hardware, data and control stacks.

Runway’s own numbers suggest the release is meant to be practical, not just demonstrative. The company says a default chunked call on a single H100 server takes about 0.55 to 0.6 seconds round trip, and that each policy call returns 24 action rows at 15 Hz, or 1.6 seconds of motion. It also said fine-tuning on DROID used successful episodes totaling about 277 hours on 128 H100s for 50,000 optimizer steps, while RT-1 fine-tuning used about 78,000 episodes and 95 hours from the full RT-1 train split over about 28 hours on 32 H100s.

For now, the main takeaway is that Runway has moved Praxis-1 from a research preview into a publicly available robotics model with open weights, a permissive license and enough deployment detail for hardware startups and robotics labs to begin their own experiments.

Sources