DEV Community

Toàn Thành thy
Toàn Thành thy

Posted on

Runway Just Open-Sourced a Robot Brain: What Praxis-1's 27B World Action Model Actually Changes

Runway is best known for video generation. On November 11, the company released something that points in a very different direction: Praxis-1, a 27-billion-parameter model that controls real robots — now with open weights under the Apache 2.0 license, including for commercial use.

This is not a new model. Runway introduced Praxis-1 in September and committed then to shipping it as an open model. The news is that the weights, the inference code, and the fine-tuning code are now public. Here is what the release actually contains, what the numbers mean, and where the honest limits are.

What a "world action model" is

Most robot policies map observations directly to motor commands. Praxis-1 takes a different route: it first predicts what the world will look like after the next action — a short video of the future — and then reads that imagined future to decide what motion to output.

The model is a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together. It is built on the same large-scale video pretraining behind Runway's general world models. The bet is that internet video already teaches most of what a robot needs to know about how objects behave and how tasks unfold, and that robot-specific data is only needed to ground that knowledge in a particular body.

Runway says its research has shown that web video matches teleoperated robot video as pretraining data for policy performance. If that holds up, the bottleneck for a capable robot policy shifts from collecting expensive robot demonstrations to how much general video a model can learn from. That is a big claim, and it is the one to watch as independent groups start testing the release.

The benchmark numbers

On NVIDIA's RoboLab-120 leaderboard, Praxis-1 reaches 46.2% success across 1,200 episodes. Runway reports it as the top-performing world action model, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8%, and π0.5 at 28.0%, and says it outperforms every open-weight VLA.

Read that number carefully. 46.2% is a first-place finish, and it is also a model that fails more than half of its episodes. RoboLab-120 is a simulated benchmark, and simulation success is not the same as reliable behavior on a factory floor. The honest reading is that open-weight robot models are improving fast, and that none of them is close to production-grade reliability yet.

The more interesting number is about generalization. Fine-tuned on Google's RT-1 robot — which never appeared in Praxis-1's training data — the model reaches 80.0% success on SimplerEnv, well above the original RT-1 policy at 52.4%. A single checkpoint has also adapted to different embodiments with light fine-tuning: launch partners Noble Machines, Standard Bots, and Ultra have run it on a legged humanoid, a six-axis industrial arm, and a dual-arm warehouse robot.

What you actually get

The release covers the weights and the inference and fine-tuning code, with support for both full fine-tuning and LoRA. That last part matters most for small teams: LoRA lets you adapt the model to your own embodiment, cameras, and tasks without retraining the full model.

It is available on GitHub, on Hugging Face with native LeRobot support, and in vLLM-Omni, with hosted inference on Reactor. Runway says it runs out of the box across NVIDIA hardware, from data center GPUs like H100 and B200, to workstation GPUs like RTX 5090 and RTX PRO 6000, to Jetson Thor on the robot itself.

Why this release matters

Three things make it worth paying attention to, beyond the leaderboard:

1. The data bottleneck moved. If general video is a valid substitute for teleoperated robot data at pretraining time, the hardest part of building a robot policy stops being data collection and becomes compute and adaptation. That favors teams with GPUs and a specific robot over teams with a fleet of teleoperators.

2. Adaptation is cheap by design. LoRA fine-tuning to a new embodiment means a lab with one robot can plausibly take a frontier-scale world model and make it theirs, instead of training a small policy from scratch.

3. Open weights change who can verify the claims. Every number in this post comes from Runway's own announcement. With the weights public, independent groups can now reproduce the RoboLab-120 results, test the sim-to-real gap, and find out whether the video-pretraining story survives contact with other people's hardware.

The limits, stated plainly

  • 46.2% on a simulated benchmark is progress, not reliability. A robot that fails half its tasks is a research demo.
  • A 27B model still needs serious hardware. "Runs on a workstation GPU" is true and still means a high-end card.
  • Open weights do not automatically produce safe robot motion. A model that can imagine futures and act on them needs the same safety thinking as any other deployed agent — arguably more, because its outputs move physical things.
  • All performance figures so far are Runway-reported. Treat them as a starting point for your own evaluation, not as settled fact.

Praxis-1 is not the moment robots get solved. It is the moment the strongest open approach to robot control becomes something anyone can download, inspect, and try to break. That is a better place for the field to be than a leaderboard screenshot — and the next few months of independent testing will tell us how much of Runway's story holds up.

Source: Runway's official announcement, "Praxis-1 Is Now Available" (November 11, 2026).

Top comments (1)

Collapse
 
alexshev profile image
Alex Shev •

The interesting test for a world-action model is not only whether it can plan a sequence but whether it knows when its state estimate is uncertain. Publishing failure modes, recovery behavior, and the gap between simulation and physical feedback would make the capability claim much more useful to builders. In robotics, calibrated uncertainty is often more valuable than one impressive demo.