Runway opens 27B Praxis-1 world-action model under Apache 2.0

By Rebecca Stone

Runway released Praxis-1, a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together, on Wednesday, Nov. 11, and reported a 46.2% success rate on Nvidia’s 1,200-episode RoboLab-120 benchmark.

Runway, best known for AI video generation, released Praxis-1’s weights under Apache 2.0, a license that permits commercial use, and included inference and fine-tuning code. First announced in September, Praxis-1 supports LoRA so developers can adapt it to their own robots, cameras and tasks, and builds on the same video pretraining as Runway’s general world models.

What does the open release put in developers’ hands?

Weights sit on GitHub and Hugging Face, with support for the LeRobot library and vLLM-Omni. Runway says Praxis-1 runs on Nvidia hardware ranging from H100 and B200 data-center GPUs to RTX 5090 workstation cards and the Jetson Thor robot computer.

That combination links an open checkpoint to two software stacks and three classes of Nvidia hardware. It is the commercial pitch: developers can start with Runway’s model and adapt it to a chosen robot setup. “We view open world models as a compounding advantage,” said Andy Chen, Runway’s head of robotics.

Weights were only part of the package. Runway also supplied the inference and fine-tuning code, plus LoRA support. That matters for teams working with different cameras, robots and tasks. The caveat is equally concrete: the performance numbers presented at launch come from Runway’s own testing.

How does Praxis-1 compare with competing robot models?

On Nvidia’s RoboLab-120 benchmark, Runway reports that Praxis-1 succeeded in 46.2% of 1,200 episodes. The comparison field included:

• FLUX 3 Action: 42.9% success
• Nvidia’s Cosmos3 Nano: 36.8% success
• Physical Intelligence’s π0.5: 28.0% success

Runway calls Praxis-1 the top world action model on the leaderboard and says it beats every open-weight vision-language-action model. The Robot Report noted that it ranks fifth overall. Those statements address different scopes: a world-action-model category and an overall table. The launch material does not reconcile the two rankings.

Runway also reported a transfer result after fine-tuning Praxis-1 on Google’s RT-1 robot, which the company says was absent from its training data. On SimplerEnv, the adapted model scored 80.0%, compared with 52.4% for the original RT-1 policy. The 27.6-point gap is the release’s clearest numerical comparison for adapting a video-and-action model to an unfamiliar robot platform.

Who is testing Praxis-1 in physical warehouses?

Three launch partners provide the first named physical deployments:

• Noble Machines builds a legged humanoid for heavy industrial handling.
• Standard Bots makes six-axis arms for machine tending and palletizing.
• Ultra deploys dual-arm stations that pack warehouse orders.

The warehouse test is narrower. In a Noble Machines task, Runway says Praxis-1 learned from half an hour of examples to place wooden planks in a container 15 times out of 15. It cleared an unfamiliar mixed pile in 6 of 10 attempts.

Those are Runway’s task results, not a broad measure of warehouse reliability, but they connect benchmark scoring to a defined manipulation job and a specific robotics customer.

How crowded is the world-model race?

Runway is entering a field with substantial model and capital backing. Nvidia has opened four Cosmos 3 models for robot simulation, while Nvidia and Unity backed Emulate’s $700 million round for world models.

Praxis-1 arrives with a different route to adoption: Apache 2.0 weights, inference and fine-tuning code, and named hardware support from data-center GPUs down to a robot-specific computer.

Open weights may widen experimentation, yet Praxis-1 still faces Nvidia’s expanding simulation toolkit and a well-funded Emulate effort. The next commercial test is whether developers can reproduce Runway’s reported transfer gains beyond its benchmark tasks and launch partners, turning an open release into repeatable robot deployments as competition intensifies.