Runway Opens Praxis-1 Weights Under Apache 2.0

Runway on 11 November 2026 said Praxis-1 is now available “to everyone with open weights under the Apache 2.0 license.” The company, which first introduced the model in September, describes Praxis-1 as “our state-of-the-art world action model” and as a “27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together.” Weights, inference code and fine-tuning code — full fine-tunes and LoRA — are included, and commercial use is allowed under that licence.

The official checkpoints are on Hugging Face, including runwayml/praxis-1-base, plus embodiment-specific variants such as praxis-1-droid and praxis-1-yam and LeRobot layouts of the same weights. Those repositories were created between 3 and 10 November and carry an Apache 2.0 licence; the public availability announcement is the 11 November news post. The base model card describes a 27.2-billion-parameter mixture-of-experts video-diffusion policy with a 0.3-billion-parameter action expert, bf16 weights occupying 54.4 GB, and 15 Hz control in 24-step action chunks. It is also explicit about a practical limit: the base checkpoint “has no per-embodiment action heads, so it does not run as a policy as is.” Users are told to add the heads for their rig and fine-tune from the reference code in runwayml/praxis-1-oss. Frozen dependencies named on the card, the Wan2.2 VAE and a Qwen 3.5 2B text encoder, are distributed separately under their own licences.

Runway also points developers to GitHub, Hugging Face with LeRobot support, vLLM-Omni, and hosted inference on its Reactor service, and says the stack can run from H100-class servers down to NVIDIA Jetson Thor. Launch partners named in the news post are Noble Machines, Standard Bots and Ultra.

Runway’s own robot scores, and what they are not

The performance case is Runway’s. The company says Praxis-1 is “the top-performing world action model on NVIDIA’s RoboLab-120 leaderboard,” with “46.2% success across 1,200 episodes, compared with 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5.” After fine-tuning on Google’s RT-1 robot, it reports “80.0% success on SimplerEnv,” “above the original RT-1 policy at 52.4%.” Those numbers are developer-benchmark results published by Runway. They are not an independent safety evaluation, and they are not proof that a robot running the model is safe to deploy.

The open-weights release still changes what researchers and robotics teams can actually do. IFM’s PAN 2 was a recent open world model that also shipped training code; Praxis-1 is the action-conditioned relative of that idea, predicting video and controls in one set of weights. Ai2’s ToyBlocks took the other open-robotics path, shipping a world and a dataset for learning. Runway is offering a large pretrained policy that, once heads are added, can be fine-tuned on a specific arm or mobile rig, including under a commercial Apache 2.0 licence rather than a research-only term.

For a company already in the video-generation business, the release also extends an existing developer channel. Runway’s Dev API had already been adding third-party video models; Praxis-1 is a robotics-facing checkpoint on Hugging Face and GitHub, not just another endpoint on that video API. The two can sit side by side without being the same product.

Open weights are not a safety certificate

The honest reading is narrower than “a general robot brain is now free to download.” What shipped on 11 November is a large open checkpoint, Apache 2.0 code for inference and fine-tuning, and a set of Runway-reported scores on RoboLab-120 and a SimplerEnv RT-1 fine-tune. The base weights, on Runway’s own model card, still need embodiment-specific heads and a fine-tune before they act as a policy. The leaderboard lead is a company claim. None of that is a substitute for task-level safety testing on the robot that will actually run the model, in the environment where it will run.

That is the useful caution, not a reason to ignore the release. Open world-action weights with commercial licence terms are still rare at this size, and the combination of a 27-billion-parameter mixture-of-experts trunk, public fine-tune code and named launch partners is the news. The facts to watch are independent RoboLab or real-robot replications, and whether the base-checkpoint caveat stays visible once third-party fine-tunes start to circulate. Until those arrive, Praxis-1 is an open world action model with a commercial licence, and Runway’s benchmark lead should be quoted as Runway’s.

Sources

Runway, “Praxis-1 is now available,” 11 November 2026. https://runway.com/news/company-news/praxis-1-is-now-available
Hugging Face, runwayml/praxis-1-base model card and Apache 2.0 licence (repositories created 3–10 November 2026; public release announced 11 November). https://huggingface.co/runwayml/praxis-1-base