Runway releases open-weight Praxis-1 robot action model
Runway says Praxis-1, a model for predicting video and robot actions together, is now available with open weights under the Apache 2.0 license. The company describes it as a 27-billion-parameter mixture-of-experts diffusion transformer.
Key takeaways · 4
- 01
Praxis-1 is available with open weights under Apache 2.0, including for commercial use, according to Runway.
- 02
The model combines video prediction with robot-action prediction and takes camera, instruction and robot-position inputs.
- 03
Runway reports a 46.2% RoboLab-120 success rate; the results have not been independently verified.
- 04
Runway says LoRA can adapt Praxis-1 without retraining the full model, but a new robot system still needs additional training.
A model for video and robot actions
Runway describes Praxis-1 as a state-of-the-art world action model and its first world action model, introduced in September and built on large-scale video pretraining.[1] The company says the model predicts video and robot actions together.[1] The Decoder reports that Praxis-1 takes camera images, text instructions and robot-position information as inputs, and adjusts its control as new camera frames arrive.[2] That setup makes ongoing camera input part of the control loop, rather than relying only on an instruction supplied at the start.[2]
Access, license and hardware
Runway says Praxis-1 is available on GitHub and Hugging Face, in vLLM-Omni, and through hosted inference on Reactor.[1] It says the files needed to run and adapt the model on users’ own hardware are released under Apache 2.0, including for commercial use.[1] Runway lists NVIDIA data-center, workstation and Jetson Thor hardware as supported.[1] The Decoder reports that the standard model needs a GPU with 80 GB of memory.[2] It also reports Runway estimates a compressed version could fit on an RTX 5090 with 32 GB of VRAM, but peak memory use has not been measured.[2]
Benchmark results and caveats
Runway reports a 46.2% success rate across 1,200 RoboLab-120 episodes.[1] It reports scores of 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5 on the same benchmark.[1] Runway says Praxis-1 leads NVIDIA’s RoboLab-120 leaderboard among world action models and outperforms every open-weight VLA.[1] The Decoder reports that Runway conducted the RoboLab-120 tests itself and that the results had not been independently verified.[2] It also reports that in ten attempts per task, Praxis-1 succeeded 88 times and FLUX 3 Action succeeded 80.[2]
Adapting the model to robots
Runway says a single Praxis-1 checkpoint has been adapted to different robot types with light fine-tuning, and that LoRA can adapt it without retraining the full model.[1] For a Google RT-1 robot that was absent from the model’s training data, Runway reports 80.0% success on SimplerEnv after fine-tuning, compared with 52.4% for the original RT-1 policy.[1] The Decoder reports that adapting Praxis-1 to a Franka arm used about 277 hours of DROID recordings, while training material for a dual-arm robot ranged from 1,400 to 2,600 hours.[2] It also reports limitations: similar-looking objects can cause confusion, vague instructions can be difficult, and longer tasks require individual steps to be specified.[2]
For teams evaluating robot-control models, Praxis-1 offers an openly licensed option to test and adapt, including for commercial use. The reported benchmarks are a starting point, not a substitute for validation on a team’s own robots and tasks.
Why it matters
AI training for your team
kju.ai is built for teams: daily sessions, team analytics, and rollout support for your organization.
How this developed
14 November 2026
Runway releases open-weight Praxis-1 robot action model
