Primary use stated or inferred cautiously from official documentation.
Praxis-1 SO-101
Runway video-and-action policy adapted to LeRobot’s SO-100 / SO-101 follower arm. The checkpoint appeared on November 12 and the official code added its configuration that day. It controls joints and the gripper; it is a distinct adaptation from DROID and YAM.
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
Limits and integration
| API ID | runwayml/praxis-1-so101 |
|---|---|
| Model type | Robotics · Video · Multimodal |
| Access model | Downloadable weights with acceptance of conditions |
| License | Apache-2.0 · dependencies under their own licenses |
| Deployment | Self-hosted |
| Release | 12 NOV 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image · Robot state |
| Exit | Video · Robot actions |
| Context window | Instruction, one or two cameras and joint state |
| maximum output | 24 steps at 15 Hz; six-dimensional actions |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | Not published |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · LeRobot |
Manipulation with an SO-100 or SO-101 arm and external and wrist cameras. Each action contains five joint angles relative to the start of the chunk and absolute gripper closure between zero and one. It predicts video and actions in 1.6-second chunks.
The BF16 reference requires an 80 GB GPU and downloads 54.4 GB of weights; file size is not total runtime memory. File access requires signing in and agreeing to share data on Hugging Face. Wan2.2 VAE and Qwen3.5-2B are obtained separately and retain their own licenses.
What it can do
Orientation
Manipulation with an SO-100 or SO-101 arm and external and wrist cameras. Each action contains five joint angles relative to the start of the chunk and absolute gripper closure between zero and one. It predicts video and actions in 1.6-second chunks.
Context
Instruction, one or two cameras and joint state · 24 steps at 15 Hz; six-dimensional actions
Tools and integration
Not published
Access
Hugging Face · Runway code
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | Not published | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Not published | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No hosted API tariff is verified. Self-hosting consumes infrastructure. NVFP4 is a QAT variant of the same checkpoint, not another catalog model: it downloads the full BF16 weights and packs experts at load time. Experimental inference requires an SM120 GPU and does not remove dependencies.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
How to read the results
No numerical SO-101 score is published and verified in the reviewed model cards. RoboLab-120’s 46.2% result belongs to DROID and is not attributed to this adaptation. Before controlling a real robot, validate cameras, units, joint limits and safe stopping.
i We have not found published benchmarks for this model.
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
Chronology
runwayml/praxis-1-so101
SO-100 / SO-101-specific checkpoint; the NVFP4 variant retains its own runtime configuration.