Runway Praxis-1 open weights: 27B world action model for robots released under Apache 2.0 with DROID, YAM and NVFP4 checkpoints
Runway released Praxis-1, the robot control model it introduced on Sept. 30, 2026, as open weights under the Apache 2.0 license on Nov. 11, 2026. The release covers the weights plus inference and fine-tuning code, including for commercial use. Runway calls Praxis-1 a world action model: a 27.2-billion-parameter mixture-of-experts diffusion transformer that predicts the next stretch of video and 1.6 seconds of robot actions together. NVIDIA's RoboLab-120 simulation leaderboard lists it at 46.2%, or 555 of 1,200 episodes, fifth overall and first among world action models, ahead of Black Forest Labs' FLUX 3 Action at 42.9%, NVIDIA's Cosmos3-Nano-Policy at 36.8% and Physical Intelligence's π0.5 at 28.0%. Runway says launch partners Noble Machines, Standard Bots and Ultra have run it on their own hardware: a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. None of the sources describes a paid customer deployment.
Source: Runway: Praxis-1 is now available (Nov. 11, 2026)
Servospace take: the robot model is free, the compute is not
Servospace analysis: Apache 2.0 removes the license fee. A robot maker or warehouse integrator can fine-tune Praxis-1, ship it inside a commercial product and keep its changes private without paying Runway. Hugging Face lists eight Praxis-1 model repositories, all tagged Apache 2.0: base, DROID single-arm and YAM bimanual checkpoints, LeRobot versions of each, and 4-bit NVFP4 versions of the DROID and YAM checkpoints. Each holds about 27.2 billion parameters, and the DROID checkpoint takes 54.4 GB, per Hugging Face repository metadata checked at 03:17 UTC on Nov. 12, when the eight repositories showed 268 downloads in total.
The lock-in moves to hardware. Runway's GitHub quick start asks for one CUDA GPU with at least 80 GB of memory, and NVIDIA's leaderboard lists Praxis-1 at about 70 GB of VRAM, against 40 GB for NVIDIA's own Cosmos3-Nano-Policy and more than 8 GB for π0.5. Every target Runway names is NVIDIA hardware, from H100 and B200 data center GPUs to Jetson Thor, which NVIDIA lists with 128 GB of memory. NVIDIA's Jetson AGX Thor developer kit starts at $3,499. Runway's README says the 4-bit version shrinks the policy from 54.4 GB to 17.5 GB. With the text encoder (4.6 GB) and the VAE (1.4 GB) that is about 23.5 GB before activations, so a 32 GB RTX 5090 should fit, though peak memory has not been measured, and a 24 GB GPU does not.
Who makes money: NVIDIA on the chips, and Reactor, a separate company that hosts Praxis-1 at $1.80 per hour on its price list and sells post-training. At that price, one robot calling it around the clock costs about $15,800 a year, Servospace math. Runway's own cost comparison is company-run and covers three tasks: it put Praxis-1 at $0.04 per completed task on a $3.95-per-hour H100, against $1.87 for GPT-6 Astra. Runway has not disclosed Praxis-1 revenue; its launch post asks teams planning a deployment to contact its robotics team.
More on Nvidia, Physical Intelligence, Hugging Face and robot ai.
Sources
- Runway: Praxis-1 is now available (Nov. 11, 2026)
- Runway Research: Praxis-1, From Video Pretraining to Robot Control (technical report, Nov. 11, 2026)
- Runway Research: Introducing Praxis-1 (Sept. 30, 2026)
- Runway Robotics: Praxis-1 product page
- GitHub: runwayml/praxis-1-oss repository and README (repo license Apache-2.0)
- Hugging Face: runwayml Praxis-1 collection
- NVIDIA Research: RoboLab-120 leaderboard
- Reactor: Praxis-1 hosted inference and pricing
- Reactor docs: Praxis-1 variants and session price
- NVIDIA Newsroom: Jetson Thor now available (Aug. 25, 2025)
- Neural Brew: Runway open-sources Praxis-1 under Apache 2.0 (Nov. 11, 2026)
- The Robot Report: Runway introduces Praxis-1 world action model (Oct. 2, 2026)
New to robots? Read our guide: How robots work: a beginner's guide.
Summarized by Servospace from the sources above, drafted with AI help and checked against those sources before publishing. Claims are the companies' or outlets' unless we say otherwise. How we work.