# Is Praxis-1 SOTA?

Pending benchmarks. Announced Nov 11, 2026; no independent results yet.

As of Nov 22, 2026, Praxis-1 is pending benchmarks on isitsota: announced Nov 11, 2026, with no independent results on the boards we track yet.

Runway's open world action model for robot control: a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together. Announced as publicly available on November 11, 2026, after a September teaser. Official pages describe a video backbone plus a smaller action expert, with 27.2B parameters in total and about 2.8B active per token in the backbone. Apache 2.0 weights and inference and fine-tuning code are public. No first-party hosted API with a published model id is stated.

Canonical: https://isitsota.com/m/praxis-1
Last updated: 2026-11-11
Tracking since: 2026-11-11

## Announced facts

| Fact | Value |
| --- | --- |
| Lab | Runway |
| Announced | Nov 11, 2026 |
| API | Not yet |
| Open weights | Published · Public, ungated Hugging Face repos with sharded safetensors (runwayml/praxis-1-droid, praxis-1-yam and praxis-1-base). Inference, a policy server, a MuJoCo simulation, and full and LoRA fine-tuning code are in github.com/runwayml/praxis-1-oss |
| Parameters | 27.2B total; about 2.8B active in the backbone |
| Type | World action model for robot control. A video backbone predicts the next scene; an action expert of about 314M parameters reads those features and outputs 1.6 seconds of actions (24 rows at 15 Hz) per call (Runway tech report) |
| Architecture | Mixture-of-experts diffusion transformer in the latent space of a frozen Wan2.2 VAE, with a frozen Qwen3.5-2B text encoder. 17 of 20 feed-forward layers are MoE (one shared expert plus 128 routed experts, eight of them used per token); the action expert is a dense 16-layer transformer (Runway tech report) |
| Checkpoints | Three official Apache 2.0 checkpoints on Hugging Face: runwayml/praxis-1-droid (fine-tuned on DROID), runwayml/praxis-1-yam (fine-tuned on the bimanual YAM) and runwayml/praxis-1-base (the mid-trained base). Official pages also mention LeRobot-format repos, vLLM-Omni, and planned NVFP4 checkpoints (Runway launch post and tech report) |
| License | Apache License 2.0 for the weights and the inference and fine-tuning code, including commercial use (Runway launch post, Hugging Face LICENSE files, and GitHub) |
| Hosted access | No first-party API model id is stated. The official launch post points to hosted inference on Reactor and a try-it path on Runway Dev, plus NeMo Automodel tutorials for post-training (Runway launch post) |
| Hardware | Official launch post: runs on NVIDIA data-center GPUs such as H100 and B200, workstation GPUs such as RTX 5090 and RTX PRO 6000, and Jetson Thor on the robot. The tech report says the 27.2B bf16 checkpoint is 54.4 GB and runs on one 80 GB GPU |
| Related | Separate from Runway Gen 4 / Gen 4.5 video and image models, and from GR00T and π0.5 |

Tags marked unsure are not stated on the lab's own pages.

## What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

- [x] Publicly announced with a source
- [x] Usable by the public (API or weights)
- [ ] Listed on an independent board we track
- [ ] Enough results for a rank in a category
- [ ] Holds across two snapshots (crowns only)

## What the lab claims

Self-reported. Shown for context; not used in any verdict.

- Lab claim: self-reported 46.2% success on NVIDIA RoboLab-120 (555 of 1,200 episodes under the official 10-trials-per-task protocol), described by Runway as the top world action model and the top open-source robot foundation model on that board as of publication, ahead of FLUX 3 Action at 42.9%, Cosmos3 Nano at 36.8%, and π0.5 at 28.0%, and ahead of open-weight VLAs including π0.5 and GR00T N1.6 (Runway launch post and tech report).
- Lab claim: self-reported as finishing the same three YAM tasks 15.6 times faster than GPT-6 Astra on average, at 51 times lower cost per completed task (Runway tech report).
- Lab claim: self-reported 80.0% success on SimplerEnv after fine-tuning on Google's RT-1 robot, which never appeared in its training data, compared with 52.4% for the original RT-1 policy (Runway launch post and tech report).

## Sources

- LAB: [Praxis-1 Is Now Available (Runway News)](https://runway.com/news/company-news/praxis-1-is-now-available) (2026-11-11)
- PAPER: [Praxis-1: From Video Pretraining to Robot Control (Runway technical report)](https://runway.com/research/praxis-1-video-pretraining-to-robot-control) (2026-11-11)
- REPO: [runwayml/praxis-1-droid on Hugging Face](https://huggingface.co/runwayml/praxis-1-droid)
- REPO: [runwayml/praxis-1-oss on GitHub](https://github.com/runwayml/praxis-1-oss)
- LAB: [Runway on X: Praxis-1](https://x.com/runwayml/status/2120533081182577116) (2026-11-11)
- LAB: [Runway on X: Praxis-1 technical report](https://x.com/runwayml/status/2120533108017791460) (2026-11-11)

Tracking continues until an independent board lists it.

## Related

- Lab: [Runway](https://isitsota.com/labs/runway.md)
- All pending releases: https://isitsota.com/releases.md
- JSON: https://api.isitsota.com/v1/models/praxis-1
