On November 11 Runway released Praxis-1, its world action model, with open weights plus inference and fine-tuning code (full and LoRA) under Apache 2.0, including commercial use. It is a 27B-parameter mixture-of-experts diffusion transformer (about 2.8B active per token) that predicts future video and robot actions together, returning 1.6 seconds of motion per call at roughly 0.55 seconds round trip on one H100. Runway reports 46.2% success on NVIDIA RoboLab-120, first among world action models and open robot foundation models, ahead of FLUX 3 Action (42.9%) and pi0.5 (28.0%). On three tasks at the same station it finished 15.6x faster than GPT-6 Astra driving the arm, at 51x lower cost per completed task. Weights are on Hugging Face with LeRobot and vLLM-Omni support.
Key Takeaways
- ✓Size: 27.2B params total, 26.9B backbone with about 2.8B active per token plus a 314M action expert; ~54.4 GB in bf16, fits one 80 GB GPU
- ✓RoboLab-120 (1,200 episodes): 46.2% vs FLUX 3 Action 42.9%, Cosmos3 Nano 36.8%, pi0.5 28.0% (vendor figures)
- ✓Real Franka FR3, ten tasks: 88/100 vs FLUX 3 Action 80/100; fine-tuned on RT-1 it hits 80.0% on SimplerEnv (original RT-1: 52.4%)
- ✓Versus GPT-6 Astra driving the same arm: 15.6x faster, $0.04 vs $1.87 per completed task (51x cheaper)
- ✓NVFP4 quantization shrinks weights to 17.5 GB; RoboLab sample 46.5% vs 47.8% in bf16
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the industry pain point
Robotics keeps hitting the same wall: real robot data is scarce and teleoperation is expensive. Runway introduced Praxis-1 in September and on November 11 delivered on its promise to open it. The bet is that general video is the most scalable pretraining source, so the model learns how objects and hands move before it learns control.
Architecture and mechanism
Praxis-1 has a 27B video backbone (17 MoE layers, each with 128 routed experts picking 8 plus one shared expert), a roughly 314M-parameter action expert, and thin per-robot projections. Working in 1.6-second chunks, it predicts the next video segment and 24 action rows at 15 Hz together; the action expert reads the backbone's imagined future to infer the motion. Training runs in four stages: video pretraining, embodied-video mid-training, joint video-and-action mid-training, and per-robot fine-tuning.
Benchmarks (vendor figures)
RoboLab-120: 46.2% (555 of 1,200 episodes), first among world action models and open robot foundation models; the top two entries are frontier-model agents (86.0% and 57.1%) that are far slower. Real Franka, ten tasks: 88/100 vs FLUX 3 Action 80/100. Fine-tuned on just 29 minutes of teleop data, it succeeded in all 15 simulated trials on a Noble Machines robot.
Getting started
Hugging Face hosts DROID, bimanual YAM and base checkpoints plus NVFP4 variants; the GitHub repo runwayml/praxis-1-oss ships a policy server, a MuJoCo simulation, and full and LoRA fine-tuning code. It runs on H100, B200, RTX 5090 and Jetson Thor. Runway notes that look-alike objects cause most failures, multi-step tasks work best with one instruction per step, and guidance settings must be retuned per robot.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.