Robotics model announced in September is now public
Video-generation AI company Runway on Nov. 11 (Nov. 12 Japan time) made Praxis-1, an AI model that generates robot actions, available to all users. It is an open-weight model whose weights are published, under the Apache 2.0 license. Runway calls it a "world action model": building on the large-scale pretraining it has used for video generation, it predicts a robot's next movements together with the video frames in between.
Runway announced Praxis-1 in September. At the time, the model was being tested with partners, and Runway said a public release would come in the coming months. It named 3 partners, Noble Machines, Standard Bots and Ultra, each running the model on its own robots.
According to the latest announcement, the 3 companies have used it on a legged humanoid robot, a six-axis industrial arm and a two-armed warehouse robot. Runway has now made the same model available to all robotics developers and researchers.
Weights and code allow commercial use; video and text models carry separate licenses
Praxis-1 is published on GitHub and Hugging Face and supports Hugging Face's robotics library "LeRobot" and the inference software "vLLM-Omni." Hosted inference through Reactor is also available. Runway said everything needed to run and adapt the model on one's own hardware is released under Apache 2.0, including for commercial use. That covers the weights and the inference and fine-tuning code, which supports both full retraining of the model and "LoRA," which trains only part of it.
However, running Praxis-1 requires 2 published models that are used frozen, without further training. They are the VAE from "Wan2.2," which compresses video, and "Qwen3.5-2B," which reads instructions. The GitHub README states that each of the 2 has its own license, and the Hugging Face model card also says the 2 are distributed separately and are subject to their respective licenses.
On Hugging Face, Runway listed "praxis-1-base," a base model that serves as the starting point for fine-tuning, along with models tuned for the one-armed "DROID" and the two-armed "YAM," among others.
46.2% success rate on NVIDIA benchmark, ahead of 3 models Runway compared
Praxis-1 is a 27 billion (27B) parameter model that predicts video and actions together, and it uses a "MoE" design that selects some of several expert components. According to Runway, it posted the highest score among world action models on the leaderboard of NVIDIA's "RoboLab-120" benchmark and outperformed all open-weight vision-language-action (VLA) models. Runway said its success rate across 1,200 trials was 46.2%, higher than "FLUX 3 Action" at 42.9%, "Cosmos3 Nano" at 36.8% and "π0.5" at 28.0%.
When fine-tuned for Google's "RT-1" robot, which was not in its training data, its success rate on the "SimplerEnv" benchmark was 80.0%. RT-1's original control model scored 52.4%, according to Runway. All of these figures come from Runway's announcement, and no third-party verification could be found as of Nov. 12.
Hardware: announcement lists RTX 5090, model card specifies 80GB GPU
On supported hardware, the official news post says the model runs out of the box on NVIDIA hardware ranging from data center GPUs such as the H100 and B200 to workstation GPUs such as the RTX 5090 and RTX PRO 6000, as well as "Jetson Thor," which is installed on robots.
The Hugging Face card for the DROID model, on the other hand, says inference requires one 80GB GPU. According to the card, the model has 27.2B (27.2 billion) parameters and its weights are 54.4GB.
The GitHub README also calls for a GPU with at least 80GB of memory in the default configuration. For GPUs of the RTX 5090 generation, it provides an experimental configuration that compresses part of the model to 4 bits, reducing the weights to about 23.5GB. The README says this should fit on the 32GB RTX 5090 but that peak usage has not yet been measured. It says this configuration changes the robot action output and has not yet been tested in repeated real action loops on RTX 5090-generation GPUs, and it asks users to measure accuracy before using it.
The model card also states that on part of Runway's own evaluation data, the DROID model's generated video fully followed instructions 52.5% of the time (95% confidence interval: 35.8% to 68.3%). This is a separate evaluation from the RoboLab-120 figures in the news post.
Google offers through partners, while Runway highlights NVIDIA collaboration
Around the same time, Google DeepMind announced "Gemini Robotics 2.5" for robots on Nov. 10 (Nov. 11 Japan time). Google is starting with partners in the robotics field and has not disclosed an API or pricing.
Runway published the Praxis-1 weights in a form anyone can download. As its reason, it said that for the U.S. to lead in AI that operates in the physical world, American models need to be open and interoperable. Runway is a founding member of the "Cosmos Coalition" and said it will work with NVIDIA to open the technology foundation to developers. It also provided a guide to fine-tuning with NVIDIA's training software "NeMo Automodel." It directs companies planning deployments to contact its robotics division.
Runway has continued to make developer-focused announcements centered on video generation, including making its developer API platform "Runway Dev" available in Claude Code on Nov. 4 (Nov. 5 Japan time). Praxis-1 can also be tried on Runway Dev.
