Runway — the company that made AI video generation mainstream — released a 27-billion-parameter robot control model on November 11. It is called Praxis-1, it runs on NVIDIA hardware, and it is available on Hugging Face under Apache 2.0. That means commercial use is included and there is no licensing agreement to negotiate. A video AI company now ships what currently leads the open-weight robotics leaderboard.
Not a Vision-Language-Action Model #
Most robot foundation models map sensor inputs directly to motor commands. Praxis-1 does something different. It first predicts what the world will look like after the robot moves, then uses that imagined future to decide which motion to output. Runway calls this a world action model, and the distinction is not marketing — it reflects a specific architectural choice with real consequences.
The backbone is a 26.9-billion-parameter mixture-of-experts diffusion transformer, the same video pretraining technology behind Runway’s general video models. A separate 314-million-parameter action expert reads the backbone’s internal representations through cross-attention, then converts them into robot movements. Crucially, the backbone never reads the action expert. This separation keeps the two systems from interfering with each other during training.
The researchers found that most remaining errors come from wrong video predictions, not from the action module. If the model imagines the wrong future, it moves the robot incorrectly. Fix the video prediction and the robot behavior improves automatically. The full architecture is documented in Runway’s technical paper.
The Benchmark Story #
Praxis-1 leads NVIDIA’s RoboLab-120 open-weight leaderboard at 46.2% success across 1,200 simulated episodes, ahead of FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%), and π0.5 (28.0%). When fine-tuned on Google’s RT-1 robot data and tested on SimplerEnv, it reaches 80% success, compared to the original RT-1 policy at 52.4%.
What 46.2% actually means: the model fails more than half its tasks in simulation. That makes this a strong research result, not a production system. If you build retry logic into your deployment — two attempts per task — effective success rates climb to roughly 71%. Design your applications accordingly.
| Model | RoboLab-120 Success | License |
|---|---|---|
| Praxis-1 (Runway) | 46.2% | Apache 2.0 |
| FLUX 3 Action | 42.9% | Closed | | Cosmos3 Nano | 36.8% | Apache 2.0 | | π0.5 | 28.0% | Apache 2.0 |
What You Can Download Today #
Runway published three checkpoints on Hugging Face alongside inference code, full fine-tuning code, and LoRA support:
-
Base mid-trained model — use this if you plan to fine-tune from scratch on your own robot data
-
DROID fine-tune — the right starting point for single-arm tabletop manipulation; this is the checkpoint that produced the 46.2% leaderboard result
-
YAM fine-tune — optimized for bimanual (dual-arm) robot configurations The model also integrates with LeRobot and vLLM-Omni. Runway’s Reactor service offers hosted inference for teams that do not want to manage GPU infrastructure. Full details are on Runway’s release page.
Hardware Requirements #
The full bf16 weights come in at 54.4GB, which requires an 80GB GPU — an H100, B200, or RTX PRO 6000. Runway has a quantized NVFP4 variant planned at 17.5GB, which would fit on a 24GB RTX 5090. Edge inference on NVIDIA Jetson Thor is also supported. At cloud H100 rates, running the model costs approximately $0.12 per 10-minute task.
What You Need Before You Start #
Open weights do not mean plug-and-play. Before running Praxis-1 on actual hardware, there are several steps you should not skip:
- Map your camera observations and joint state to the model’s expected input format
- Choose the correct action head for your gripper or manipulator type
- Collect demonstration data specific to your tasks — DROID used approximately 277 hours of successful episodes, but Runway reports that simple tasks can be learned from 30 minutes of examples
- Test in simulation before any hardware contact
- Build retry logic from day one; a 50% success rate on first attempt is not a bug in your setup
A detailed developer checklist covers the full setup sequence including supply-chain verification and embodiment-specific validation steps.
Why Runway Open-Sourced This #
Runway’s Andy Chen put the efficiency case plainly: “Praxis-1, built with more compute and data, learned it from half an hour of examples and got it right 15 times out of 15.” The company’s argument is that video pretraining generalizes. Because the backbone already understands how objects move in the world, the robot-specific fine-tuning can be relatively lightweight.
The strategic bet is also straightforward. Open weights invite the community to build fine-tunes for new embodiments, which extends Runway’s reach without Runway having to do the work. For the broader field, the release shifts the resource constraint from data collection — which requires expensive teleoperator fleets — toward compute and adaptation, which favors teams with GPU access. Whether that benefits the community or simply advantages well-funded labs is a question worth watching as the first community fine-tunes appear.