Runway opens Praxis-1, its 27-billion-parameter robot model, free for commercial use
Runway made Praxis-1 generally downloadable on November 11, releasing the weights and fine-tuning code for its robot model under Apache 2.0 — including for commercial use. Runway reports 46.2% success across 1,200 episodes on Nvidia's RoboLab-120 benchmark, ahead of FLUX 3 Action at 42.9%, but the benchmarks are Runway's own and 46.2% still means failing more than half of the simulated tabletop tasks.
📊 Scorecard: our Bullish call on $NVDA is scored from $173.73, the last close before it ran. Its first result comes after the next close.
The News
Runway made Praxis-1 generally downloadable on November 11, releasing the weights and the fine-tuning code for its robot model under the Apache 2.0 licence — including for commercial use. Praxis-1 is a 27-billion-parameter 'world-action' model: a mixture-of-experts diffusion transformer that predicts video and robot actions together, trained on large-scale video. The idea is that a model that has watched vast amounts of video of the physical world can learn how objects move and how hands interact with them, then turn that knowledge into actions a robot can take. Runway is a private company, so the public read-through for investors is Nvidia — Praxis-1 runs out of the box on Nvidia hardware, from H100 and B200 data-centre chips to the Jetson Thor robotics computer.
Runway reports 46.2% success across 1,200 episodes on Nvidia's RoboLab-120 benchmark, a set of simulated tabletop manipulation tasks, for the DROID fine-tuned checkpoint, ahead of FLUX 3 Action at 42.9%. When fine-tuned for Google's RT-1 robot it scored 80.0% on SimplerEnv, a simulation benchmark, versus 52.4% for the original RT-1 policy. The release includes three checkpoints — the base model, a version fine-tuned on the DROID robot dataset, and a bimanual version fine-tuned on YAM, for two-handed tasks — plus inference code and full and LoRA fine-tuning code on GitHub and Hugging Face, with support for LeRobot, the open-source robotics framework. Runway's research write-up lists the three checkpoints (https://runway.com/research/praxis-1-video-pretraining-to-robot-control).
Two caveats matter. First, this is not a brand-new model: Praxis-1 was first introduced in September, and the November 11 event is the public open-weight release. Second, the benchmark figures are Runway's own, with no independent replication published yet — and 46.2% on RoboLab-120 still means the model fails more than half of the simulated tabletop episodes. Open weights are a real gift to robotics researchers, but the headline number is a long way from a robot that works reliably.
Praxis-1 is free to download — and every lab that runs it will do so on Nvidia chips.
Why It Matters
Open-weight releases change who can build on a model. When weights are free under Apache 2.0, any lab or startup can download Praxis-1, fine-tune it on its own robot and data, and ship a product without paying Runway or asking permission. That is how open models spread: they become the default starting point, and the ecosystem — tools, datasets, tutorials — grows around them. For robotics, where data is scarce and every lab trains on its own hardware, a shared starting point matters more than in most fields.
The benchmark lead is modest but real. Beating FLUX 3 Action by 3.3 points on RoboLab-120 and raising the RT-1 policy's SimplerEnv score from 52.4% to 80.0% when fine-tuned for its robot suggests the video-first training approach works — the model has learned something about the physical world from watching video that transfers to controlling a robot arm. But simulation is not the real world: a 46.2% success rate in a simulator, where lighting and physics are controlled, will be lower on a real robot in a messy kitchen.
For Nvidia, the read-through is straightforward. Praxis-1 is designed to run on Nvidia hardware, and every lab that downloads it and fine-tunes it will do so on Nvidia chips. Open robot models are a demand engine for GPUs in the same way open language models were: they put more workloads on the hardware. The model itself is free; the compute to run it is not.
Analyst's Take
Our call is Bullish. An open-weight robot model that beats its nearest rival on a standard benchmark, with three checkpoints and full fine-tuning code, is a genuine accelerant for robotics research — and it runs on Nvidia hardware, which is where the money is for public-market investors. The September introduction means this is a release, not a debut, but the release is the event that matters: weights in the wild beat weights on a slide.
What to watch: independent replication of the RoboLab-120 numbers, and whether labs start publishing results built on Praxis-1 rather than on their own models. If the model becomes the default base for robot manipulation research, Runway has set a standard — and Nvidia sells the chips that run it. If the benchmarks do not hold up outside Runway's lab, the release fades into the pile of open models that nobody builds on.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What is Praxis-1?
Praxis-1 is Runway's robot model — a 27-billion-parameter AI that predicts video and robot actions together, so it can watch a scene and decide what a robot arm should do next. It was first introduced in September 2026.
What does 'open weights' mean?
It means Runway has released the trained model files for anyone to download and use, under the Apache 2.0 licence — including for commercial products — along with the code to fine-tune it on new robots and tasks.
How good is it?
Runway reports 46.2% success across 1,200 episodes on Nvidia's RoboLab-120 benchmark, ahead of FLUX 3 Action at 42.9%. But 46.2% still means failing more than half of the simulated tabletop tasks, and the figures are Runway's own — no independent lab has replicated them yet.
What is a mixture-of-experts model?
It is a model built from many smaller specialist sub-models, called experts, with a router that sends each input to only a few of them. The model can be very large in total while only using a fraction of its size for any one decision, which keeps it fast.
Why is Nvidia the ticker if Runway is private?
Runway is not listed on a stock exchange, so there is no Runway stock to buy. Nvidia is the public read-through: Praxis-1 runs out of the box on Nvidia hardware, from H100 and B200 chips to the Jetson Thor robotics computer, so wider adoption means more demand for Nvidia chips.
👁
Source: Runway