Runway released Praxis-1’s model weights and inference and fine-tuning code on November 11, 2026, under Apache 2.0. The public release gives robotics researchers access to the world action model introduced in September, when access was limited to early partners.
What changed on November 11
The development is a release milestone for an existing model. Runway’s September introduction described partner testing ahead of a future public launch. Its November 11 announcement says the weights and inference and fine-tuning code are now available under Apache 2.0, including for commercial use, with support for full fine-tuning and low-rank adaptation, or LoRA.
Runway describes Praxis-1 as a 27-billion-parameter model that predicts video and robot actions together. For university robotics groups, the immediate change is access to an open weights model they can run and adapt locally, rather than a promise of a later download.
What researchers can download
The release separates a starting model from policies adapted to particular robot configurations. Runway’s Hugging Face collection lists the base model, DROID and YAM fine-tunes, alongside LeRobot versions. These artifacts serve different research purposes:
- Base checkpoint: Runway’s technical report identifies a mid-trained model, providing a starting point for further adaptation.
- Robot-specific checkpoints: The repository describes DROID as a single-arm configuration with exterior and wrist cameras, and YAM as a bimanual configuration with a top camera and two wrist cameras.
- Adaptation and evaluation code: The repository links fine-tuning instructions and simulation evaluation documentation. The technical report also identifies a policy server and a MuJoCo simulation among the released components.
This distinction matters when comparing experiments: a base checkpoint and a robot-specific fine-tune represent different starting conditions. A research record or model card should identify which artifact was used, rather than recording only the Praxis-1 name.
What the release means for universities
Local weights and adaptation code expand the experiments a university team can conduct on its own robot. Computing requirements still depend on the execution path: the repository’s standard quick start specifies Python 3.12, a CUDA 13.0 driver and a CUDA GPU with at least 80 GB of memory. That documented configuration is a concrete consideration for research computing teams assessing an initial experiment.
For research offices supporting a reproducible AI experiment, the release makes artifact selection especially important. Recording the checkpoint, code revision, robot and camera configuration, and evaluation protocol would help distinguish changes to the model from changes to the experimental setup. This is a reporting implication of the release, rather than evidence that results transfer between laboratories.
Performance claims and reported limitations
Runway reports a 46.2% success rate across 1,200 episodes on RoboLab-120. That figure is a company-reported benchmark result; it does not establish performance or safe deployment on a university’s robot.
The technical report identifies errors involving look-alike objects, difficulty with vague instructions, and problems chaining multiple task steps from a single prompt. These limitations give researchers specific failure modes to investigate alongside task success. Public availability changes access to the model; the report’s limitations remain relevant to interpreting its behavior.
Sources
- Runway: Praxis-1 Is Now Available — November 11, 2026; checked 2026-11-14.
- Runway Research: Introducing Praxis-1 — September 2026; checked 2026-11-14.
- Runway: Praxis-1 repository and README — checked 2026-11-14.
- Runway: Praxis-1 model collection — checked 2026-11-14.
- Runway Research: Praxis-1: From Video Pretraining to Robot Control — November 2026; checked 2026-11-14.
