Runway Praxis-1 Is Now Open Weights — a 27B World Action Model for Robots

Runway has made Praxis-1 open weights. The 27-billion-parameter "world action model" for robots — a system that predicts video and robot actions together — is now available under the Apache 2.0 licence, including for commercial use, with weights plus inference and fine-tuning code. Announced on 11 November 2026, the release puts one of the most capable robot models ever built into the hands of anyone with a powerful enough GPU. Runway says it scored 46.2% on NVIDIA's RoboLab-120 benchmark, the best result among world action models and open-weight models. That number is genuinely impressive — and it still means the model fails more than half the time.

What Praxis-1 Actually Is

Praxis-1 is a 27B-parameter model — 27.2 billion parameters in total, with about 2.8 billion active per token — built as a mixture-of-experts diffusion transformer. Those terms need unpacking. A "mixture-of-experts" model is one that routes each input to a subset of specialised sub-networks, so only part of the model runs at any moment; that is why only 2.8 billion of the 27.2 billion parameters are active per token. A "diffusion transformer" is an architecture borrowed from video generation, which builds its output step by step from noise. And a "world action model" is the idea that ties it together: instead of treating vision and movement as separate problems, Praxis-1 predicts video and actions together, learning how the world looks and how a robot should move through it in a single model.

What Is Actually Open

Open weights means the trained model is available to download and use, and the Apache 2.0 licence is one of the most permissive in open source: it allows commercial use, modification and redistribution. Runway is releasing the weights plus inference and fine-tuning code, including both full fine-tuning and LoRA — a lighter method that adapts a model without retraining all of it. Three checkpoints are available: a base mid-trained model, one fine-tuned on DROID, a large dataset of robot demonstrations, and one fine-tuned for bimanual YAM work, meaning two-armed manipulation.

The model is available on GitHub and Hugging Face, with native support for LeRobot, the popular open-source robotics framework from Hugging Face. It also runs through vLLM-Omni and through hosted inference on Reactor, Runway's own platform. That spread matters: a researcher can download the weights and fine-tune them, a startup can run them on its own hardware, and a team without a GPU cluster can call the hosted version instead. Open weights is only useful if the model is actually runnable, and Runway has covered the main ways people will want to run it.

AI-generated illustration by INORYVA.

The Benchmark: 46.2% and What It Means

The headline number is 46.2%: Praxis-1's success rate on RoboLab-120, NVIDIA's robot benchmark, where the DROID-tuned checkpoint succeeded in 555 out of 1,200 episodes. The benchmark is 120 tabletop manipulation tasks, each tried ten times. Runway says that makes Praxis-1 the top world action model and the top open-weight VLA on the RoboLab-120 leaderboard — and fifth overall. That last detail is the one to hold on to: fifth overall means four other models scored higher, and "number one" is only true inside the two categories Runway is claiming.

Model	RoboLab-120 success rate
Runway Praxis-1 (DROID-tuned)	46.2% (555/1,200)
FLUX 3 Action	42.9%
Cosmos3 Nano	36.8%
π0.5	28.0%

But 46.2% also needs to be read plainly. It means the model fails more than half of the tasks it is given, in a controlled simulation, on tabletop tasks that are deliberately simple: pick up a cup, move a block, stack an object. These are not factory jobs, warehouse shifts or household chores. They are the easiest version of physical work, and the best open model in the world still fails most of them. RoboLab-120 is a simulated benchmark, and simulation flatters models: the lighting is perfect, the objects are known, and nothing unexpected happens. Real-world performance will be worse, not better. Anyone describing 46.2% as factory-ready is misreading the number.

Why Open Weights Matters for Robotics

Robotics has lagged behind language AI in openness. The best language models have open-weight competitors within months of release; the best robot models have mostly stayed behind closed doors, trained on proprietary data and run on proprietary hardware. That has slowed the whole field: researchers cannot build on what they cannot see, and startups cannot experiment with models they cannot afford to licence.

Praxis-1 changes that arithmetic. A 27B model with published weights, permissive licensing and LeRobot support is the kind of release that creates an ecosystem: fine-tunes for new robots, new tasks and new environments, all shared back. The open-source robotics community has been waiting for a foundation model it can actually use, and the comparison with the open-source language world is instructive. Our Impeccable GitHub skill and Hallmark GitHub skill coverage shows how quickly high-quality open releases attract contributors — and robotics, with its hardware costs and safety risks, needs that shared effort more than most fields.

AI-generated illustration by INORYVA.

What Praxis-1 Cannot Do Yet

Honesty requires saying what the release does not solve. The model fails most of its benchmark tasks. It needs an 80 GB GPU, or a quantised version that trades some accuracy for size. It was trained largely on video and simulation, and the gap between simulation and the real world — the "sim-to-real" gap — is the oldest unsolved problem in robotics. A robot that stacks blocks in a simulator can still drop them in a kitchen, because kitchens have glare, clutter, pets and people.

AI-generated illustration by INORYVA.

What to Watch Next

The test of this release is not the announcement but what happens in the next few months. Watch for independent reproductions of the 46.2% figure — if other labs can match it, the number is real; if they cannot, it was a favourable run. Watch for fine-tunes: the first community releases for new robots and new tasks will show whether Praxis-1 generalises or only works on the hardware it was trained for. And watch the leaderboard: fifth overall is a strong start, but the models ahead of it are not standing still, and the next release from any of them could take the top spot back.

The most important watchpoint is the simplest one: real robots, doing real work, outside a lab. Noble Machines, Standard Bots and Ultra are the first test of that. If their robots get visibly better with Praxis-1 inside them, the open-weights bet will have paid off in the way that matters. Until then, Praxis-1 is a strong benchmark result and a generous release — and the most honest description of it is the one Runway's own numbers give: the best open robot model yet, and still failing more than half the time.

Key Takeaways

• Runway's Praxis-1 is now open weights under Apache 2.0, including commercial use, with weights plus inference and fine-tuning code (full and LoRA) on GitHub and Hugging Face.
• Praxis-1 is a 27B-parameter (27.2B total, ~2.8B active per token) mixture-of-experts diffusion transformer "world action model" that predicts video and actions together.
• On NVIDIA's RoboLab-120 it scored 46.2% (555/1,200 episodes) — top among world action models and open-weight VLAs, but 5th overall on the leaderboard.
• It beats FLUX 3 Action (42.9%), Cosmos3 Nano (36.8%) and π0.5 (28.0%) — but 46.2% still means failing more than half its tasks, in simulation.
• RoboLab-120 is a simulated benchmark (120 tabletop tasks x 10 trials); the figures are Runway's own, with no independent replication yet.
• Three checkpoints are available (base, DROID-fine-tuned, bimanual YAM-fine-tuned); it needs a single 80 GB GPU at 54.4 GB in bf16, or 17.5 GB in the 4-bit NVFP4 version.
• Runway also reports 80.0% on SimplerEnv with Google's unseen RT-1 robot, vs 52.4% for the original RT-1 policy.

Conclusion

Praxis-1 is the most important open release in robotics this year, and the numbers explain why: a 27-billion-parameter model, free for commercial use, beating every other open robot model on the most widely watched benchmark. But the same numbers set the limit. A model that succeeds 46.2% of the time in simulation is a research milestone, not a worker. The release matters because it gives the whole field something to build on — and the field's job now is to close the gap between 46.2% and the reliability real work demands.

Sources

Runway — Praxis-1 Is Now Available

Runway Research — Praxis-1: Video Pretraining to Robot Control

The Robot Report — Runway's Praxis-1 World Action Model for Robotics Is Now Available