Runway open-sources Praxis-1: a 27B robot model that runs on one H100, under Apache 2.0

What shipped, from Runway's own announcement (Nov 11, 2026), read first-hand: Praxis-1 is a 27B-parameter mixture-of-experts diffusion transformer, first introduced in September as Runway's first world action model, now open to every developer. Runway reports 46.2% success across 1,200 episodes on NVIDIA's RoboLab-120 benchmark, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and π0.5 at 28.0% — Runway's figures, data as of November 2026, not measured by us. The release is on GitHub, on Hugging Face with native LeRobot support, and in vLLM-Omni, with hosted inference on Reactor; launch partners Noble Machines, Standard Bots and Ultra have been running it on their own hardware, and fine-tuned on Google's RT-1 robot — which never appeared in its training data — it reaches 80.0% on SimplerEnv against the original RT-1 policy's 52.4% (Runway's figure).

The hardware maths, from Runway's research page: 27.2 billion parameters in total, 54.4 GB in BF16 — so the full model fits one 80 GB GPU — and packing the routed experts in NVFP4 shrinks the policy weights to 17.5 GB. Runway lists H100 and B200 data-centre GPUs, RTX 5090 and RTX PRO 6000 workstation cards, and Jetson Thor on the robot itself as running it out of the box. Three checkpoints are in the Hugging Face collection: the base model and DROID and bimanual YAM fine-tunes.