What happened
Runway says Praxis-1 is now available to robotics developers and researchers as an open-weight model. The company introduced it in September as its first world-action model, built on the large-scale video pretraining used for its general world models. Before the public release, launch partners Noble Machines, Standard Bots and Ultra ran it on their own hardware: respectively, a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. Runway describes Praxis-1 as a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together.
The announcement 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. Its reported success rate is 46.2% across 1,200 episodes. Runway lists FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and π0.5 at 28.0%; the company dates these figures to November 2026. These are results as presented by Runway, rather than an independently reported evaluation in the announcement.
Weights are available on GitHub and Hugging Face, with native LeRobot support. Runway also lists availability in vLLM-Omni and hosted inference on Reactor, and says developers can try the model on Runway Dev. The Apache 2.0 release covers weights, inference code and fine-tuning code, and Runway says commercial use is permitted. Fine-tuning options include full fine-tuning and LoRA. For deployment planning, the company directs teams to contact its robotics group.
Runway says Praxis-1 runs out of the box across NVIDIA hardware, including H100 and B200 data-center GPUs, RTX 5090 and RTX PRO 6000 workstation GPUs, and Jetson Thor on the robot. The company’s stated adaptation approach starts from video pretraining: it says the model learns how objects behave and tasks unfold from large-scale video, then can be fine-tuned for a new embodiment or domain. Runway reports that a single checkpoint adapted to multiple embodiments with light fine-tuning alongside partners, and that LoRA can adapt it to a team’s embodiment, cameras and tasks without retraining the full model.
The announcement also gives a transfer result for a fine-tuned Google RT-1 robot, which Runway says was absent from Praxis-1’s training data. On SimplerEnv, the adapted model reportedly reached 80.0% success, compared with 52.4% for the original RT-1 policy. Runway says its research found web video matched teleoperated robot video as pretraining data for policy performance. The announcement frames open weights as a way for hardware developers to run, inspect and adapt the model on their own terms.
What the archive adds
New Runtime’s earlier coverage of Verne’s V0.5 release on November 5 described another open-weight world-action model for industrial robots, this time paired with a fine-tuning platform. That earlier Field Note makes Praxis-1 a new release in a developing category, not a restatement: Runway emphasizes a broader range of robot embodiments, its benchmark results, and a cross-hardware adaptation path.
Trend signal
This release does not yet fit a pattern New Runtime tracks. It raises practical questions about model portability and robot runtime, but the announcement does not establish a durable interoperability layer or a general agent architecture. Its reported benchmark results and supported hardware are useful signals to watch as other robotics developers test whether open weights translate into repeatable performance across embodiments.
Why now
Runway says it committed to releasing Praxis-1 as an open model when it announced the model in September. The current release fulfills that commitment by publishing the weights and implementation code, rather than merely announcing a future opening. The company connects the move to its work as a founding member of the Cosmos Coalition and to a stated goal of advancing open world models for physical AI.
Why it matters
For robotics teams, the release changes what can be tested without starting from a closed vendor interface: the weights, inference path and fine-tuning code are available under a license Runway says permits commercial use. That gives a team room to inspect behavior, try local execution on supported NVIDIA hardware, and adapt with LoRA before committing to a full fine-tune. The cited results offer a starting benchmark, not proof that the model will work safely or reliably in a particular workspace.
The most consequential claim is that video pretraining can provide a useful head start across embodiments, reducing the amount of robot-specific data needed for adaptation. If that holds in a team’s setting, it could make iteration across cameras, manipulators and tasks less dependent on collecting a separate large dataset for every configuration. Teams still need to validate real-world task completion, failure handling and safety on their own hardware; a leaderboard score and a reported simulation result do not settle those deployment questions.
Limits and uncertainty
The performance figures and claims about transfer and data efficiency come from Runway’s announcement. The announcement does not provide enough detail to assess RoboLab-120’s task composition, statistical uncertainty, evaluation protocol or independent replication. Nor does the reported SimplerEnv result establish performance on physical robots across environments. The release lists supported NVIDIA hardware, but gives no inference latency, memory requirements, throughput or deployment cost. Teams should also verify that the released code and dependencies meet their operational needs before relying on the model.
Reader takeaway
Praxis-1 is a substantial open release for robotics builders: the weights, inference code and adaptation tools are available, with a reported route from general video pretraining to multiple robot embodiments. Treat Runway’s benchmark and transfer numbers as claims to test, not deployment guarantees. If you work on robot policies, the practical next step is to reproduce the published setup and measure adaptation on your own hardware and tasks.