Black Forest Labs held the open-weight robotics crown for seven weeks. On Friday, Runway took it.
Praxis-1, Runway's world action model for robot control, is now downloadable with weights and source code under the Apache 2.0 license. On Nvidia's RoboLab-120 benchmark it scores 46.2 percent, ahead of FLUX 3 Action at 42.92 percent. The gap is small. What sits behind it is not.
FLUX 3 Action arrived on September 25 as a 7-billion-parameter World Action Model, open weights, first place on RoboLab, integrated natively into Hugging Face's LeRobot stack with Jetson edge deployment recipes built alongside Nvidia. It looked like Black Forest Labs was planting a flag in physical AI with the same playbook that made FLUX the default image model: small, open, fast, everywhere.
Runway answered with a model roughly four times larger, a permissive license, and a benchmark sheet that puts it on top of the World Action Model category.
What Runway actually shipped
Praxis-1 was not a surprise drop. Runway introduced it in late September as a partner-only preview. Since then, Standard Bots, Noble Machines and Ultra have run it on a humanoid, an industrial arm and a dual-arm warehouse robot respectively.
The mechanism is the part worth understanding. Praxis-1 takes camera frames, a text instruction and the robot's joint state, predicts how the scene will change if a given action is taken, then inverts that prediction into motor commands. It re-plans on every incoming frame. Where a conventional policy learns a direct mapping from pixels to actions, this one simulates the immediate future first and picks the motion that lands there.
That approach leans on video pretraining. Runway started from a video model that already understands how objects move, then fine-tuned on human action footage and robot teleoperation data, betting that general video knowledge can replace some of the expensive, robot-specific datasets that normally gate this work.
- License: Apache 2.0 for weights and training code, gated download on Hugging Face, code on GitHub
- Scale: reported at 27 billion parameters, against 7 billion for FLUX 3 Action
- Hardware: about 80 GB of GPU memory for the standard checkpoint, with a compressed variant estimated to fit a 32 GB RTX 5090
- Pre-adapted builds: shipped for several robot platforms, while the base model still needs extra training for a new one
The benchmark fight is closer than the headline
Runway's own numbers are the whole case, and Runway ran every test itself. Nothing in that report has been independently verified, which matters because the World Action Model category is young and the leaderboard is small.
- RoboLab-120 success rate: Praxis-1 at 46.2 percent, FLUX 3 Action at 42.92 percent
- Overall leaderboard: Praxis-1 sits fifth, behind two AI agents and two robot models built on different architectures
- Real Franka arm, ten tasks with ten attempts each: Praxis-1 succeeded 88 times, FLUX 3 Action 80 times
- Unseen robot transfer: fine-tuned on about 95 hours from a mobile platform Google used for RT-1, Praxis-1 averaged 80 percent in simulation against 52.4 percent for RT-1
The FLUX 3 Action result is genuine, too. It took first place at launch by beating OASIS WAM at 39.0 percent and Nvidia Cosmos3-Nano at 36.8 percent, while using 56 percent fewer parameters than the previous best open model and running up to 3.95 times faster.
Which is the awkward part. FLUX 3 Action loses by 3.3 points on accuracy, but at a quarter of the parameter count and with an efficiency story that matters once you try to run a policy on a battery-powered machine. On real hardware the margin is eight successes out of a hundred attempts. That is a lead, not a rout.
Why generative media labs build robot brains now
Both models are bets on the same thesis: the best robot controller is a video model wearing a different hat. Predicting the next frames of a video and predicting the next frames of a robot camera feed are nearly the same problem, and converting predicted motion into joint commands is cheap compared with learning control from scratch.
Runway is best known for Gen-4.5. Black Forest Labs is best known for FLUX. Neither was a robotics company two years ago. Both now ship robot policies, leaning on the pretraining advantage they built in image and video generation. Black Forest Labs' Blattmann made the ambition explicit in an interview this week, describing image generation as a detour and robotics as the destination. Runway's release reads like a company that reached the same conclusion and decided to arrive first with a license nobody has to negotiate over.
Runway's data numbers show how much of the old problem remains. Adapting Praxis-1 to a Franka arm took roughly 277 hours of recordings from the public DROID dataset. A dual-arm platform needed between 1,400 and 2,600 hours. Video pretraining lowers the floor. It does not remove the dataset.
Runway's report is also candid about failures. Praxis-1 confuses similar-looking objects, vague instructions trip it up, and long tasks have to be decomposed into explicit steps because the model cannot reliably plan a multi-stage sequence alone. Runway frames it as a fast execution layer for a general reasoning model, not an autonomous agent.
That framing is honest and also a hedge. Every open-weight robotics release this year has the same shape: strong short-horizon manipulation, weak planning, and a benchmark score that depends on who ran the test.
FLUX 3 Action has the stronger deployment story today, with LeRobot integration, Jetson recipes and 7 billion parameters that fit comfortably on edge hardware. Praxis-1 has the better scoreboard and a license that asks nothing of you. The next honest data point will be an independent RoboLab run, and the one after that will be whichever model a factory keeps after the demo.
For now, the open-weight robot race has a new leader on paper, and the company that spent two years teaching the world how to generate images is one benchmark behind.
Comments