Runway Praxis-1: Open-Weight 27B Robot Model Ships Under Apache 2.0
Praxis-1 is a 27B-parameter mixture-of-experts robot control model from Runway, released with open weights on 11 November 2026 under Apache 2.0, which allows commercial use. It predicts video and robot actions together. The default build needs one 80 GB GPU such as an H100, while an NVFP4 option shrinks the policy to about 17.5 GB for Blackwell cards.
On this page
What Runway released on 11 November
Runway first showed Praxis-1 in September with access limited to a few launch partners. On 11 November 2026 it made the model generally available, publishing the weights along with inference and fine-tuning code under the Apache 2.0 license, which permits commercial use.
Praxis-1 is what Runway calls a world action model. It is a 27B-parameter mixture-of-experts diffusion transformer that predicts future video frames and robot actions at the same time, building on the company's video generation research. The Hugging Face model card describes a 27.2B trunk paired with a 0.3B action expert, trained on robot and human video covering 18 different embodiments, and running control at 15 Hz in chunks of 24 action steps.
The release comes as several checkpoints. There is a base model for adapting to new hardware, a fine-tune for the single-arm DROID setup, and a fine-tune for the bimanual YAM arms, each also packaged in the LeRobot format. The model can be run from GitHub, Hugging Face, vLLM-Omni, or hosted inference through Reactor.
How Praxis-1 scores against other robot policies
Runway's headline number comes from NVIDIA's RoboLab-120 leaderboard, where the DROID-tuned Praxis-1 reports 46.2% success across 1,200 episodes. Runway lists FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8%, and pi0.5 at 28.0% for comparison. The Robot Report notes that this places Praxis-1 fifth overall on that leaderboard, while Runway frames it as the top world action model and the top open, end-to-end policy there.
A second test checks transfer to hardware the model never saw. After fine-tuning on Google's RT-1 robot, Praxis-1 reached 80.0% on SimplerEnv, against 52.4% for the original RT-1 policy. Runway's partner examples are smaller but concrete: on a warehouse plank task the model learned from about half an hour of demonstrations and succeeded 15 out of 15 times, and it cleared an unseen messy pile in 6 of 10 attempts.
These figures come from Runway and its partners. Independent replication on the open weights is the next thing to watch.
The hardware you need to run it yourself
For local builders the memory budget is the real story. The repository says the default bf16 policy checkpoint is about 54.4 GB, plus roughly 6 GB for the text encoder and video VAE, so it wants one GPU with at least 80 GB of memory, such as an H100.
The smaller path is NVFP4. It quantizes only the routed experts, cutting them from 51.3 GB to 14.4 GB and the full policy to about 17.5 GB, or around 23.5 GB of weights in total before activations. That route works on Blackwell-class GPUs, including the RTX 5090, but not on an H100, and the docs say a 24 GB card does not fit while a 32 GB RTX 5090 should but has not been measured yet. Runway separately lists Jetson Thor for running the model onboard a robot.
Quantization also carries a measured cost. The repository reports one RoboLab run with little change (279 of 600 against 287 of 600) and a harder suite with a clear drop (138 of 352 against 188 of 352), and advises measuring before deployment.
Why an open robot policy matters for private AI
Most capable robot policies are still reached through a vendor's cloud. An Apache 2.0 model that can be downloaded, inspected, and adapted lets a lab or factory keep its camera feeds and task data on its own machines. Runway's head of robotics, Andy Chen, told The Robot Report that open world models give hardware developers flexibility and control they do not have today.
Fine-tuning remains heavy. Adapting the DROID model with LoRA needs one 80 GB GPU, and a full fine-tune of the base model is shown on eight GPUs. Note too that the weights are gated behind accepting conditions on Hugging Face, and the bundled Wan2.2 VAE and Qwen3.5 text encoder carry their own licenses. Check those before building a product on top.