What Runway released
Runway made Praxis-1 public with open weights under the Apache 2.0 licence on 11 November 2026. The licence covers the weights and Runway's inference and fine-tuning code, and commercial use is allowed. Runway calls Praxis-1 a "world action model". A model of this kind predicts the next stretch of camera video and the robot movements that go with that video, both in the same pass. Runway first showed Praxis-1 in September and tested it with early partners. At that stage The Robot Report noted that a public release was promised "in the coming months". The release page names Noble Machines, Standard Bots and Ultra as launch partners.
The Hugging Face collection holds a base checkpoint and fine-tunes for single-arm DROID rigs, the two-arm YAM setup and SO-100/SO-101 arms. It also has copies in LeRobot layout, 4-bit variants and an evaluation dataset. You can also reach the model through vLLM-Omni, the hosted Reactor service and Runway Dev, according to the release page.
How the model works
Runway's technical report puts the full model at 27.2B parameters. Nearly all of those parameters sit in a 26.9B video backbone built as a mixture of experts (MoE). An MoE layer holds many expert sub-networks and a router sends each token to a few of them, so only about 2.8B parameters do work for any one token. Seventeen of the backbone's 20 feed-forward layers use MoE. A separate dense "action expert" with 16 layers and about 314M parameters turns the backbone's output into motor commands. Praxis-1 is a diffusion transformer, meaning it builds its output by removing noise step by step. The repository recommends two denoising steps per call. Each call returns 24 rows of actions, which covers 1.6 seconds of motion at 15 Hz.
Runway's bet is that ordinary video can stand in for robot demonstrations, which are expensive to record. CTO Kamil Sindi told The Robot Report that "Most robot policies are bottlenecked by robot data, which is scarce and expensive to collect." The technical report shows the training mix that follows from that view. In the embodied-video stage, 58% of the data was egocentric human video, 20% was general video, 18% was real robot video and under 1% was gameplay. The DROID fine-tune then used about 277 hours of successful DROID robot episodes.
Praxis-1 also depends on two frozen models: the Wan2.2 video autoencoder and Qwen3.5-2B as the text encoder. The README says both come under their own licences, so check those licences as well as Apache 2.0 before you ship a product.
Benchmarks, as Runway reports them
On NVIDIA's RoboLab-120 leaderboard, Runway reports a 46.2% success rate over 1,200 episodes. The comparison models scored lower: FLUX 3 Action 42.9%, Cosmos3 Nano 36.8% and Physical Intelligence's π0.5 28.0%. Runway also fine-tuned Praxis-1 for Google's RT-1 robot, which wasn't in its training data. On the SimplerEnv benchmark that fine-tune reached 80.0% success, against 52.4% for the original RT-1 policy. All of these figures are Runway's own. AI Weekly points out that nobody has reproduced them independently yet. Its suggested checks are a third-party RoboLab-120 run and outside latency tests on Jetson Thor.
Hardware, latency and cost
The default bf16 checkpoint is about 54.4 GB, split into 17 shards. The README asks for one CUDA GPU with at least 80 GB of memory and says a single H100 is enough. On an H100 with synthetic inputs, a call after the first takes about 501 ms at the median with video guidance at 4.0, or about 344 ms at 1.0. The first call in a process takes about 35 seconds while kernels compile. In RoboLab runs against a hosted server, client-side latency was 1.1 to 1.4 seconds per call.
NVFP4 is NVIDIA's 4-bit floating-point format for Blackwell-generation GPUs. Using NVFP4 for the routed experts shrinks the policy from 54.4 GB to 17.5 GB, and total weights come to about 23.5 GB before activations. The README expects a 32 GB RTX 5090 to fit but hasn't measured one, and says a 24 GB card won't fit. NVFP4 runs on B200/B300, RTX 5090-class cards and Jetson Thor. An H100 rejects NVFP4 and stays on bf16. The SO-101 NVFP4 model card still marks the format as experimental. The card also notes that the repository is gated and that you download the full 54.4 GB whichever precision you plan to run.
The README's accuracy results for NVFP4 are mixed. On RoboLab, an NVFP4 fine-tune scored 279 out of 600 against 287 for bf16. On the harder suite, an NVFP4 fine-tune scored 138 out of 352 against 188 for the bf16 checkpoint it started from. The README adds that NVFP4 changes the actions the model outputs and hasn't yet been tested in closed loop on RTX 5090-class GPUs. Runway's advice is to measure your own checkpoint before you deploy it.
Fine-tuning costs vary with the method. LoRA (low-rank adaptation) trains small add-on weight matrices and leaves the full model untouched. A LoRA run on DROID data fits on one 80 GB GPU. Fine-tuning the base model in full for a new robot uses eight GPUs in the README's example.
Known failure modes
Runway's technical report is frank about where Praxis-1 struggles. On real robots, the most common failure is reaching for an object that looks like the target. The wrong choice usually shows up already in the video the model imagines. Vague instructions do noticeably worse than instructions that name a colour or size. Long tasks chained from one prompt are still hard, because the model has to track which steps it has finished without being told. The README lists some practical constraints too. Results change with the attention backend and the GPU. The YAM checkpoint accepts only 16:10 camera views. The text encoder has to share the policy's GPU.
What a builder could do with it
You can try Praxis-1 without owning a robot. The repository ships MuJoCo tabletop tasks that run in closed loop, with a live web view on port 8000 and MP4 output. After the installation steps in the README, the documented path is to download a checkpoint and start the simulator:
hf download runwayml/praxis-1-droid --local-dir checkpoints/praxis-1-droid
MUJOCO_GL=egl python scripts/run_sim.py --config configs/sim_droid.yamlTeams that already own an SO-100 or SO-101 arm have a ready-made checkpoint. For any other arm, the base model plus a LoRA run is the documented way to adapt Praxis-1 to your own cameras and joints. That route still needs recorded demonstrations from your own robot in LeRobot format.
For New Zealand teams, the Apache 2.0 licence matters because it has no territory or field-of-use restrictions. A startup can build a commercial manipulation product on Praxis-1 without negotiating a separate licence. Few local developers will have an 80 GB GPU at their desk, so most will rent one in the cloud. If your camera footage comes from a client's factory, packhouse or warehouse, check which region that rented GPU runs in before uploading any data. Local inference on Jetson Thor or an RTX 5090 keeps footage on site, but those NVFP4 paths are the least tested part of the release. We'd treat them as experimental until Runway or outside testers publish closed-loop results.
If you get Praxis-1 running on your own rig, you can share your results on the ASAI projects page.
Sources
- Praxis 1 is now available (Runway)
- Praxis-1: video pretraining to robot control, technical report (Runway)
- runwayml/praxis-1-oss (GitHub)
- Praxis-1 collection (Hugging Face)
- runwayml/praxis-1-so101-nvfp4 model card (Hugging Face)
- Introducing Praxis 1 (Runway)
- Runway introduces Praxis-1 world action model for robotics (The Robot Report)
- Runway open-sources Praxis-1, a 27B robot model, Apache 2.0 (AI Weekly)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.