Runway's Praxis-1 Robot Model Is Apache 2.0 and Tops NVIDIA's Leaderboard. The Winning Score Is 46.2%, and NVIDIA Is Runway's Coalition Partner.
Runway released Praxis-1 as open weights on November 11 under Apache 2.0, with inference and full and LoRA fine-tuning code, on GitHub and Hugging Face with LeRobot support, in vLLM-Omni, and hosted on Reactor. It is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and actions together. Runway says it tops NVIDIA's RoboLab-120 at 46.2% success over 1,200 episodes, against FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8%, and pi0.5 at 28.0%, and reaches 80.0% on SimplerEnv after fine-tuning on Google's RT-1 robot against 52.4% for RT-1's original policy. Every number is Runway's, the leaderboard belongs to NVIDIA, a Cosmos Coalition co-founder with a model in the table, and first place fails more than half its episodes. We read Runway's post.
By T.W. Ghost
Runway released Praxis-1 as open weights today, November 11, 2026, under Apache 2.0, which means commercial use with no territory clause and no revenue gate. The package is the weights, inference code, and fine-tuning code for both full and LoRA runs, on GitHub and Hugging Face with LeRobot support, in vLLM-Omni, and as hosted inference on Runway's Reactor. The model, introduced in September, is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together, a world action model in Runway's phrase.
The numbers are the catch, in two ways. First, their size: Runway says Praxis-1 tops NVIDIA's RoboLab-120 at 46.2% success over 1,200 episodes, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8%, and pi0.5 at 28.0%, which makes first place a model that fails more than half the time. Second, their source: every figure is Runway's, we did not read the leaderboard page, the board belongs to NVIDIA, and NVIDIA co-founded the Cosmos Coalition with Runway and has Cosmos3 Nano in the table. We read Runway's news post and research page. Here is what held up.
What Runway Shipped
| Praxis-1 open weights (Runway, November 11, 2026) | |
|---|---|
| License | Apache 2.0; commercial use |
| What ships | Weights; inference code; fine-tuning code, full and LoRA |
| Where | GitHub; Hugging Face, with LeRobot support; vLLM-Omni; hosted inference on Reactor |
| Architecture | 27B-parameter MoE diffusion transformer; predicts video and actions together |
| RoboLab-120 | 46.2% over 1,200 episodes; FLUX 3 Action 42.9%; Cosmos3 Nano 36.8%; pi0.5 28.0% (Runway's table; leaderboard not read) |
| SimplerEnv | 80.0% after fine-tuning on Google's RT-1 robot vs 52.4% for RT-1's original policy (Runway; simulation) |
| Claim | "Outperforms every open-weight VLA" (Runway) |
| Hardware | H100, B200, RTX 5090, RTX PRO 6000, Jetson Thor (Runway; no memory or speed figures) |
| Launch partners | Noble Machines, Standard Bots, Ultra |
| Affiliation | Runway is a founding member of the Cosmos Coalition with NVIDIA |
First place is 46.2%. More than half of the 1,200 episodes fail, and FLUX 3 Action is close behind at 42.9%. Leading an open-weight robot board in November 2026 means being the least unreliable entry on it.
The board and a competitor share an owner. RoboLab-120 is NVIDIA's; Cosmos3 Nano is NVIDIA's; Runway and NVIDIA co-founded the Cosmos Coalition. None of that makes the number wrong. It makes it a number to read on the leaderboard page, which we have not done yet.
Simulation is not a robot. The 80.0% is SimplerEnv, a simulator, after fine-tuning on RT-1 data. It measures how well the model adapts to a dataset, not how it performs on a floor.
Runs out of the box, size unstated. Five GPUs are named, from H100 to Jetson Thor, with no memory requirement in gigabytes and no speed number for any of them.
Verified vs Unconfirmed: The Scorecard
| Claim | Verdict |
|---|---|
| Open weights Nov 11, 2026; Apache 2.0; weights plus inference and fine-tuning code | Verified (Runway post) |
| 27B MoE diffusion transformer; video and actions together | Verified (Runway) |
| RoboLab-120: 46.2%, 42.9%, 36.8%, 28.0% | Vendor; leaderboard not read |
| "Outperforms every open-weight VLA" | Vendor claim |
| SimplerEnv 80.0% vs 52.4% | Vendor; simulation |
| Launch partners Noble Machines, Standard Bots, Ultra | Verified (Runway) |
| Runs on H100, B200, RTX 5090, RTX PRO 6000, Jetson Thor | Vendor claim; unquantified |
| Cosmos Coalition founding membership | Verified (Runway) |
Who Should Use It
Use Praxis-1 if you want a commercially usable, fine-tunable robot policy with the code to adapt it to your own arm, and you are prepared to measure it yourself.
Skip the leaderboard line until the RoboLab-120 page is read independently, and skip any plan that assumes the 80.0% carries from a simulator to a floor.
The frontier is elsewhere: the same day's contrast in open weights is MiniMax's H3-Image, released under a license that excludes the US, the EU, the UK, and Korea, against Apache 2.0 here. Our model comparison has the current standings side by side.
Sources
- Runway: Praxis-1 is now available
- Runway Research: Praxis-1, video pretraining to robot control
- GitHub: runwayml/praxis-1-oss
- Hugging Face: Runway Praxis-1 collection
*Not sure whether an open robot policy or a frontier model driving an arm fits your project? Take the free 2-minute quiz and get matched. Then read about another open-weights release where the fine print carried the story: Pika H3 Flex.*