What launched and when
Runway released Praxis-1 SO-101, a video-and-action policy fine-tuned to control LeRobot SO-100 and SO-101 follower arms. The checkpoint operates in joint space: it outputs movements for the joints and gripper, not a trajectory expressed as an end-effector pose.
The release date is November 12, 2026. The official checkpoint page confirms that the model has been published, although accessing its files requires signing in to Hugging Face and accepting its access conditions. A public repository therefore does not mean anonymous, unrestricted downloads.
The model combines a 27.2-billion-parameter video-diffusion trunk organized as a mixture of experts with a 300-million-parameter action expert. The page describes a chunk-causal architecture. These figures characterize the checkpoint; they do not directly measure accuracy, speed, or robotic task success.
The policy takes a text instruction, one or two camera views, and joint state as input. It generates 24-step action chunks at 15 Hz. Each step represents five joints and the gripper: the model expresses joint angles relative to the start of the chunk and an absolute gripper position from zero to one, where one means closed.
The checkpoint page does not publish a verified numerical score for SO-101. The 46.2% result associated with RoboLab-120 belongs to a DROID policy and does not measure Praxis-1 SO-101. The verified sources also contain no independent evaluation that would support generalizing its performance across objects, cameras, or environments.
Variants, inputs, and access
The main reference checkpoint, `runwayml/praxis-1-so101`, is a policy fine-tuned for a single SO-100/SO-101 arm. It supports exterior and wrist cameras, with one or two views. The page specifies images arranged as tiles measuring 288 by 416 pixels per view; with two views, the canvas measures 288 by 832 pixels.
The `runwayml/praxis-1-so101-nvfp4` variant starts from the same checkpoint and received 5,000 additional steps of quantization-aware training on the routed experts. It is not a smaller download: the page says the BF16 weights occupy 54.4 GB and are downloaded in full before those experts are packed at load time.
Running the experts in NVFP4 is described as an experimental mode limited to SM120 GPUs, such as the RTX 5090 and RTX PRO 6000 Blackwell. In that configuration, policy memory drops from 54.4 GB to 17.5 GB; the optional cache for packed experts occupies 14.4 GB. The lower memory use does not reduce the download size or remove the external dependencies.
Inference uses the reference code `runwayml/praxis-1-oss`. The base-version page specifies an 80 GB GPU as an inference requirement. The frozen dependencies—Wan2.2 VAE and the Qwen/Qwen3.5-2B text encoder—are distributed separately under their own licenses. The main model declares an Apache 2.0 license; that license does not replace the terms for its dependencies.
Access to the model files has a condition: Hugging Face asks users to sign in and agree to share account information to view repository contents. The pages do not publish a price for a hosted API. Self-hosting requires infrastructure, and costs depend on the hardware and service used.
Comparison of the two SO-101 checkpoints
| Aspect | Reference checkpoint | NVFP4 variant |
|---|---|---|
| Base | Praxis-1 SO-101 | Praxis-1 SO-101 |
| Downloaded weights | 54.4 GB in BF16 | 54.4 GB in BF16 |
| Described inference | 80 GB GPU | Experimental NVFP4 on an SM120 GPU |
| Stated memory use | 54.4 GB for the checkpoint | 17.5 GB with NVFP4 experts |
| Important condition | External dependencies are separate | Packs experts at load time; does not reduce the download |
What changes for users
The model turns a short instruction, visual input, and joint state into an action sequence for the arm. The page recommends direct imperative prompts, such as asking it to pick up a red cube and put it in a box. That recommendation describes the input format; it does not demonstrate that the task will succeed.
The outputs use radians for the five joints and a zero-to-one scale for the gripper. However, the page says the LeRobot follower’s readout can be configured with `use_degrees=True`: in that case, the joints are read in degrees and `gripper.pos` uses the follower’s zero-to-100 scale. An integration must keep the model’s units consistent with the robot’s.
Before testing a task, check that the configured camera matches the expected view, that images have the right size and arrangement, and that joint state is transformed correctly. It is also worth checking joint limits, gripper conversion, and emergency stopping in the system that controls the arm. Publishing a checkpoint does not by itself validate a particular physical setup.
The NVFP4 option may interest users with a compatible SM120 GPU who need to reduce memory use during inference. It does not amount to a guaranteed speedup or quality improvement: the page publishes no quantitative comparison of performance, latency, or success rates between the two configurations.
The page says that normalization statistics and the action codec are included in `config.json`; dataset statistics are not required at inference time. This helps clarify which files accompany the checkpoint, but does not replace checking the data formats and conventions of the connected robot.
Price, limitations, and scope of the release
The pages do not announce a hosted API price for this policy. Self-hosting involves downloading large weights, having enough memory, and obtaining the frozen dependencies separately. The 80 GB requirement applies to the stated inference setup for the reference checkpoint; the 17.5 GB figure applies to the experimental NVFP4 variant and should not be read as its download size.
The model generates joint-space actions and was fine-tuned for SO-100/SO-101 arms. It should not be confused with `praxis-1-droid`, a policy for DROID, or `praxis-1-yam`, intended for a bimanual YAM setup. Capabilities or metrics from those variants do not transfer to SO-101.
The page lists BF16 weights even though the variant is called NVFP4: quantization is applied to routed experts when the model loads, while other parts remain in BF16. The NVFP4 label therefore does not mean that all weights stored in the repository are compressed in that format.
The published details describe the architecture, inputs, and deployment requirements, but provide neither an evaluation score for this fine-tune nor an independent comparison. A performance figure attributed to another policy does not fill that gap. Success rates, robustness to lighting changes, and compatibility with different setups need to be evaluated under the specific conditions of use.
The checkpoint page lists an Apache 2.0 license for the main model. Wan2.2 VAE and Qwen/Qwen3.5-2B are distributed separately and have their own licenses, so using the complete system also requires checking the terms that apply to those dependencies.
What is confirmed and what should not be inferred
| Item | Scope of the checkpoint page | Does not imply |
|---|---|---|
| Checkpoint published | An official checkpoint page is accessible subject to account conditions | Anonymous downloads or access without accepting the conditions |
| 15 Hz and 24 steps | The policy’s described output configuration | A guaranteed rate on every robot |
| 17.5 GB in NVFP4 | Stated memory use when running quantized experts on compatible hardware | A 17.5 GB download or compatibility with every GPU |
| Apache 2.0 license | The license listed for the main model | That external dependencies have the same license |
| No verified SO-101 score | The pages provide no numerical metric for this fine-tune | That the policy is effective or ineffective on every task |
How to interpret the release
Praxis-1 SO-101 is a published policy that connects instructions, vision, and joint state to actions for a LeRobot arm. Its immediate usefulness depends on being able to load a 54.4 GB model, resolve external dependencies, and configure cameras and units correctly. The NVFP4 variant offers an experimental lower-memory option for specific hardware, not a smaller download or a guarantee of faster execution.
To decide whether it fits, separate three questions: whether the physical setup matches SO-100/SO-101; whether the hardware meets the memory and GPU requirements; and whether the specific task works safely and repeatably. The checkpoint pages partly address the first two, but do not provide a success evaluation that answers the third.
Runway’s official pages state that the checkpoint is available and describe its specifications. The practical conclusion is narrower: the model may be a candidate for controlled tests on the specified arm, but its results need to be checked in each setup. There is no published basis here for assigning it a general success rate or extending its figures to other robots.
Keep exploring
Corrections and transparency
If you spot incorrect or outdated information, send us a correction with the page and source we should review.
Submit a correction