Runway Open-Sources Praxis-1, Its Robot Model That Learned by Watching Video

Runway released Praxis-1 on November 11, 2026, giving away the weights, inference code and fine-tuning software for its 27-billion-parameter robot model under the Apache 2.0 licence. Anyone can download it, run it, fine-tune it and use it commercially. According to Runway, Praxis-1 is the top-performing world action model on NVIDIA's RoboLab-120 leaderboard, ahead of every other open-weight robot model.

What Happened

Runway first introduced Praxis-1 in September 2026 as its first world action model, built on the same large-scale video pretraining behind its video generation products, and promised at the time that it would ship as an open model. The November 11 release delivers on that commitment.

Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts video and robot actions together. It was trained first on general internet video, learning motion and object interaction from the vast library of how things move in the world, and then on robot-specific data that maps that knowledge onto physical actions.

The release is unusually complete for an open model. Three checkpoints are available: a base mid-trained model, a version fine-tuned on the DROID robot dataset, and one fine-tuned for the bimanual YAM robot. The weights, inference code and full and low-rank adaptation (LoRA) fine-tuning code are published under Apache 2.0, including for commercial use. It is available on GitHub and Hugging Face, in vLLM-Omni, and with hosted inference on Reactor.

On NVIDIA's RoboLab-120 leaderboard — 120 tabletop manipulation tasks, 10 trials each, 1,200 episodes in total — the DROID-tuned Praxis-1 completed 555 of 1,200 episodes, a 46.2 percent success rate. That compares with 42.9 percent for FLUX 3 Action, 36.8 percent for Cosmos3 Nano and 28.0 percent for π0.5. Runway says that is the best result among listed world action models and open-source robot foundation models.

The full model is 54.4 GB in bfloat16, fitting on a single 80 GB GPU; a 4-bit quantised version is 17.5 GB. It runs on NVIDIA hardware from H100 and B200 GPUs to Jetson Thor on the robot itself.

What a World Action Model Is

Most robot AI falls into two camps: vision-language-action models that map instructions and images directly to motor commands, and world models that predict how a scene will evolve. Praxis-1 combines the two. By predicting future video frames together with actions, the model effectively imagines the consequences of its movements before making them — a capability borrowed from video generation, where Runway has spent years learning to predict plausible futures frame by frame.

Internet video contains far more information about how the physical world works than any robot dataset, giving the model a general understanding of physics that robot data then converts into commands.

Why It Matters

Praxis-1 arrives in the middle of a crowded week for robot models. NVIDIA opened up its GR00T 2 model days earlier, and Google's Gemini Robotics 2.5 is already running on a Mercedes-Benz factory line. Runway's entry is the most permissive of the three: full weights, fine-tuning code and a licence that allows commercial use.

Two caveats belong next to the benchmark numbers. First, every figure so far is Runway's own — no independent lab has reproduced the leaderboard result yet. Second, 46.2 percent is a first-place finish that still means failing more than half the time, on simulated tabletop tasks rather than a factory floor. A model that predicts plausible actions can still crush, drop or misjudge, and open weights do not automatically produce safe robot motion. The launch partners running the model on real hardware are the right test; their results will matter more than the leaderboard.

What This Means for Founders and Investors

For founders in robotics, Praxis-1 removes a major barrier and raises a new question. The barrier was the model itself: building a capable robot foundation model from scratch was out of reach for almost every startup. The question is where value now sits. If the best robot brain is free, differentiation moves to the data, the hardware integration, the safety layer and the specific task — warehouse picking, lab automation, home assistance — where a startup can be better than a general model.

For investors, the pattern is familiar from language models: open weights commoditise the foundation layer and push value toward applications and distribution. The robotics version of that shift is now underway, and it is moving fast — three major robot model releases in a single week. The winners will be the teams that pair a good open model with a real deployment, real customers and real data from the physical world. A leaderboard score is a starting point, not a business.

Angeled.com connects founders with investors.