ProductUnited States

Runway opens Praxis-1 weights: 27B world action model scores 46.2% on RoboLab-120

Published 2 min read
Source

Runway said on Nov 11 that Praxis-1, its 27-billion-parameter world action model for robots, is available with open weights under the Apache 2.0 licence, including for commercial use. The company reports 46.2% success over 1,200 episodes on NVIDIA's RoboLab-120 benchmark against 28.0% for Physical Intelligence's pi0.5, and 80.0% on SimplerEnv after fine-tuning on Google's RT-1 robot.

Weights, inference and fine-tuning code are on GitHub and Hugging Face.

Runway, the video-model company, released Praxis-1 to all developers on Nov 11 with open weights under the Apache 2.0 licence, it said in a company post. The model was first shown on Sept 30 with three launch partners, Noble Machines, Standard Bots and Ultra, which have been running it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. The release covers the weights, inference code and fine-tuning code, with full fine-tuning and LoRA, and allows commercial use.

What Runway is giving away

Praxis-1 is a 27-billion-parameter mixture-of-experts diffusion transformer that predicts the next video frames and the next robot actions together. The GitHub repository lists single-arm (DROID rig), bimanual (two YAM arms) and SO-101 checkpoints, each with an NVFP4 quantised version; running the quick start needs one CUDA GPU with at least 80 GB of memory. Runway says the model runs on NVIDIA hardware from H100 and B200 data-centre GPUs to Jetson Thor on the robot, and is on Hugging Face with LeRobot support and in vLLM-Omni.

RoboLab-120 success rate as reported by Runway (1,200 episodes)
  • Praxis-1
    Developer
    Runway
    Success
    46.2%
    Gap to Praxis-1
    0
  • FLUX 3 Action
    Developer
    Not stated in the post
    Success
    42.9%
    Gap to Praxis-1
    -3.3 pts
  • Cosmos3 Nano
    Developer
    NVIDIA
    Success
    36.8%
    Gap to Praxis-1
    -9.4 pts
  • pi0.5
    Developer
    Physical Intelligence
    Success
    28.0%
    Gap to Praxis-1
    -18.2 pts

Scores from Runway's Nov 11, 2026 post, data as of November 2026; gaps are MEKAWIRE calculations. Simulation benchmark, not real-robot results.

Runway's own figures: 46.2% success over 1,200 episodes on RoboLab-120, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8% and pi0.5 at 28.0%. Fine-tuned on Google's RT-1 robot, which it had not seen in training, it reached 80.0% on SimplerEnv against 52.4% for the original RT-1 policy. Runway says its research found web video as useful as teleoperated robot video for pretraining.

MEKAWIRE analysis

Runway is betting that the scarce input for robot policies is video, which it has, rather than teleoperation data, which robot makers hoard.

MEKAWIRE's calculation: on Runway's numbers Praxis-1 beats pi0.5 by 18.2 points, a 1.65 times higher success rate, and the RT-1 fine-tune adds 27.6 points over the original policy. Two caveats from cross-checking. Runway calls Praxis-1 the top world action model on the leaderboard, while The Robot Report noted it ranks fifth overall, so other model types still score higher. And a 46.2% simulated success rate means more than half of the benchmark episodes still fail. The 80 GB memory floor also keeps the full model off most robot computers without the quantised checkpoints. For buyers, Apache 2.0 matters more than the score: a robot maker can now fine-tune a large video-pretrained policy without a licence negotiation. MEKAWIRE's explainer on world models in robotics covers how Praxis-1 compares with NVIDIA Cosmos and others. Signals to watch: real-robot success rates from the three launch partners, third-party fine-tunes on Hugging Face, and whether Physical Intelligence or NVIDIA answer with newer open checkpoints.

Good or bad news for the company?

 

Key facts
Model
27B-parameter mixture-of-experts diffusion transformer, predicts video and actions
Release
Nov 11, 2026, open weights, Apache 2.0, commercial use allowed
RoboLab-120
46.2% success over 1,200 episodes (Runway)
Hardware floor
one GPU with at least 80 GB for the quick start; NVFP4 versions available
Launch partners
Noble Machines, Standard Bots, Ultra
pi0.5 on RoboLab-120
28.0%, gap 18.2 points (MEKAWIRE calc.)
SimplerEnv after RT-1 fine-tune
80.0% vs 52.4% for original RT-1
Sources
MEKAWIRE Intelligence
  1. 01

    Why it matters

    Apache 2.0 weights and inference and fine-tuning code let robot developers evaluate and adapt Praxis-1 for commercial use. Runway reports 46.2% success over 1,200 RoboLab-120 episodes, an 18.2 percentage point lead over the cited pi0.5 result. This gives hardware makers another model option, but production reliability and deployment economics are not supplied.

  2. 02

    Rival analysis

    Physical Intelligence is the clearest benchmark comparator: its cited pi0.5 score is 28.0% against Praxis-1's Runway-reported 46.2%. Independent validation and details establishing comparable evaluation conditions are not supplied. NVIDIA spans both infrastructure and competing models through its physical-AI stack, while Skild AI is identified as a cross-embodiment rival without a supplied benchmark result.

  3. 03

    Valuation context

    Physical Intelligence's $5.6B figure is a disclosed valuation, alongside $1B in total funding; neither is a public market cap. Standard Bots' $263M and Ultra's $50M are funding totals, with no valuations supplied. Runway's funding and valuation, Noble Machines' investment amount and NVIDIA's public market cap are not supplied, preventing a supported valuation comparison.

  4. 04

    Related papers

    The supplied pi0.5 paper is relevant to the benchmark comparator, while SACQ addresses low-bit vision-language-action models. Magic-W0 addresses structured world-action modeling. Only titles are supplied, so their methods, results and applicability to Praxis-1 cannot be assessed.

  5. 05

    Supply-chain implications

    The quick start requires one GPU with at least 80 GB of memory, making compute capacity an evaluation constraint; NVFP4 versions are also available. Launch partners Noble Machines, Standard Bots and Ultra span humanoids, industrial arms and warehouse robots. Their integration scope, deployment volumes, component suppliers and commercial commitments are not supplied.

  6. 06

    Signals to watch

    Watch for independent RoboLab-120 replication and partner results on physical robots, including failure rates and task coverage. Seek inference latency, fine-tuning costs and measured memory requirements for NVFP4 versions. The reported 80.0% SimplerEnv result follows RT-1 fine-tuning; evidence of transfer to production settings remains missing.

  7. 07

    Analyst view

    Praxis-1 merits technical evaluation because commercial licensing and released code lower access barriers. Its reported benchmark lead supports testing, while a 46.2% success rate leaves substantial failures within that evaluation. Investment conviction should depend on repeatable partner deployments and operating costs, neither of which is supplied.

  8. 08

    Questions you should be asking

    Do Praxis-1 and pi0.5 use comparable RoboLab-120 evaluation conditions, and can independent teams reproduce the gap? What latency, compute cost and intervention rate do partners observe on physical robots? What paid commitments and deployment volumes sit behind the launch partnerships?