Runway opens Praxis-1, its 27B robot world action model, under Apache 2.0

Runway has released Praxis-1, the robot-control model it introduced to launch partners in September, as open weights under Apache 2.0, including for commercial use. The launch post went up on 11 November, and the weights, inference code and fine-tuning code are public.

What it is

Runway calls Praxis-1 a world action model. It is a 27B-parameter mixture-of-experts diffusion transformer built on the company’s video pretraining. For each policy call it predicts the next chunk of video and the next chunk of robot actions together, and the robot executes the action chunk. A small action expert reads the video model’s internal state to work out the motion, according to the tech report.

Three checkpoints are on Hugging Face: praxis-1-droid for a single arm with an exterior and a wrist camera, praxis-1-yam for two YAM arms, and praxis-1-base, the mid-trained model you adapt to your own rig. The repos are public, not gated, and tagged Apache 2.0. The code includes a policy server, a MuJoCo simulation, and full and LoRA fine-tuning. Runway also lists native LeRobot support, day-0 support in vLLM-Omni and hosted inference on Reactor.

Runway’s numbers

Runway says Praxis-1 is the top world action model on NVIDIA’s RoboLab-120 leaderboard, at 46.2% success over 1,200 episodes, against 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5. Fine-tuned on Google’s RT-1 robot, which Runway says never appeared in its training data, it reports 80.0% success on SimplerEnv against 52.4% for the original RT-1 policy. These are Runway’s own results; we have not seen independent runs yet.

The hardware fine print

The launch post says Praxis-1 “runs out of the box” on NVIDIA hardware from H100 and B200 down to RTX 5090, RTX PRO 6000 and Jetson Thor. The repo is more careful. Its quick start needs Python 3.12, one CUDA GPU with at least 80 GB of memory and the CUDA 13.0 driver, and the LoRA fine-tune also targets one 80 GB GPU. The 4-bit NVFP4 path for RTX 5090 and RTX PRO 6000 Blackwell cards is marked experimental: it changes the actions, it “has not been run in closed loop on an SM120 GPU yet”, and it is refused on H100, B200 and older cards. Runway’s report says packing the routed experts in NVFP4 cuts the policy weights from 54.4 GB to 17.5 GB. NVFP4 quantization-aware checkpoints for the DROID and YAM policies are already public on Hugging Face under Apache 2.0, but they download as 54.4 GB of bf16 weights; the 4-bit packing happens only on that experimental path. The README says the autoencoder and text encoder come from their own public weights, so check those licences separately.

Why this matters

Open robot policies have mostly been smaller VLAs, while the strongest world-model-based systems, such as Gemini Robotics 2.5, stay partner-only. A 27B world action model under Apache 2.0 gives labs something to fine-tune, inspect and ship. NVIDIA’s GR00T 2 is the obvious comparison, and an open evaluation effort like RoboColosseum is where we would expect independent numbers to appear. Until then, plan on a data-center-class GPU, not a desktop card.