RW

Praxis-1

Runway, United States 路 November 2026

World model 路 Open weights, Apache-2.0 路 Also written Praxis 1, praxis-1-droid, praxis-1-base, praxis-1-yam, praxis-1-so101

Parameters
27 B
27.2B in total according to the technical report: 26.9B in the video backbone, of which about 2.8B are active for any token,...
GPU memory
70 GB
reported by the authors
Inference latency
600 ms
single NVIDIA H100, default chunked execution, round trip per policy call; the report gives about 0.55 to 0.6 s for 1.6 s of...
Weights
Apache-2.0
Apache-2.0 for the weights and for the inference and fine-tuning code, commercial use included, according to Runway's launch post.

What it is

Praxis-1 is Runway's open-weight world action model: a 27B mixture-of-experts video diffusion transformer with a small action expert that reads the video model's internal state, so each call predicts the next 1.6 seconds of video and 24 actions at 15 Hz together. It was announced in September 2026 and released on November 11, 2026 under Apache-2.0, weights and fine-tuning code included. The DROID fine-tune solves 555 of 1,200 RoboLab-120 episodes (46.2 percent), the highest score of any open-weight model on that leaderboard, ahead of FLUX 3 Action at 42.9, Cosmos3-Nano-Policy at 36.8 and 蟺0.5 at 28.0 percent. On a real Franka it reached 88 of 100 episodes against 80 for FLUX 3 Action on the same ten tasks, in Runway's own test. The price is hardware: the quick start asks for one GPU with at least 80 GB, and a call takes about 0.55 to 0.6 s on an H100. Checkpoints exist for the DROID Franka, bimanual YAM arms and the SO-100 and SO-101.

Architecture

Backbone
Runway's video foundation model: a diffusion transformer split into a condition encoder and a velocity decoder (14 + 6 blocks), with 17 of its 20 feed-forward layers as mixture of experts, one shared and 128 routed experts per layer of which each token uses eight. It works in the latent space of a frozen Wan2.2 VAE and reads instructions through a frozen Qwen3.5-2B text encoder. No vision language model backbone.
Action head
Separate action expert: a dense 16-layer transformer of width 1024, about 314M parameters, that cross-attends into eight backbone encoder blocks. Rectified flow on video and actions, sampled in two joint Euler steps. Each call returns 24 action rows at 15 Hz (1.6 s), 21 rows on the first call of a context window. On the DROID Franka an action row is a 10-D end-effector command (position, 6-D rotation, gripper).
Action chunk
24 steps per forward pass
Control rate
15 Hz
Parameters
27.2B in total according to the technical report: 26.9B in the video backbone, of which about 2.8B are active for any token, plus about 314M in the action expert. The launch post rounds to 27B, the RoboLab leaderboard lists 27.3B.
Pretraining data
Runway's video foundation model, pretrained on general video with clips of up to 20 seconds. Then mid-training on embodied video at 15 fps, mostly egocentric human video plus robot video, general video and a little gameplay, followed by joint video and action mid-training on single arms, pairs of arms and low-cost bimanual arms with about 30 percent video-only tasks. The released DROID checkpoint was fine-tuned on about 277 hours of successful DROID episodes for 50k steps on 128 H100s.
Embodiments
Franka (DROID setup), Franka FR3 (real, Positronic station), Bimanual YAM arms, SO-100 and SO-101 arms, Google RT-1 robot (SimplerEnv), Partner robots: legged humanoid, six-axis industrial arm, dual-arm warehouse robot

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 27 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp32101 GBTraining master weights, rarely used for inference
bf16 or fp1651 GBThe usual way these checkpoints are served
int825 GBQuantised, expect some loss of precision on fine motions
int413 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 80 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference memory, reported70 GB source
Fine tuning memory, reported80 GB source
Inference latency, reported600 ms on single NVIDIA H100, default chunked execution, round trip per policy call; the report gives about 0.55 to 0.6 s for 1.6 s of motion source
Control rate15 Hz
Checkpoint download54.4 GB

The RoboLab leaderboard lists about 70 GB for inference. The README quick start asks for one CUDA GPU with at least 80 GB of memory, and the report says the 54.4 GB bf16 model runs on one 80 GB GPU. LoRA adaptation of praxis-1-droid also fits one 80 GB GPU according to the README. NVFP4 versions of the DROID, YAM and SO-101 checkpoints exist; on 600 RoboLab-120 episodes the NVFP4 model scored 46.5 percent against 47.8 percent in bf16. Runway's launch post names H100, B200, RTX 5090, RTX PRO 6000 and Jetson Thor as supported hardware without memory figures.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • SimplerEnv Google Robot, visual matching, average over 4 tasks, 864 episodes
    80%
    success rateAuthors' own evaluation after fine-tuning on RT-1 data, a robot absent from mid-training. RT-1 itself 52.4 percent; FLUX 3 Action fine-tuned by Runway on the same data 39.7 percent.source
  • RoboLab-120 overall, 10 episodes per task
    46.2%
    success rate555 of 1,200 episodes, marked as verified by the RoboLab organizers. FLUX 3 Action 42.9, Cosmos3-Nano-Policy 36.8, 蟺0.5 28.0 percent on the same board.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Real robot, Franka at Positronic (ten single-item tasks) 10 episodes per task, Runway's test of November 2, 2026
    88%
    success rateAuthors' own evaluation: 88 of 100 episodes on a Franka FR3. FLUX 3 Action 80 of 100 on the same rig.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Franka FR3 at Positronic: 88 of 100 episodes on the ten single-item tasks that FLUX 3 Action published, against 80 for FLUX 3 Action. Runway also shows bimanual YAM tasks at real speed, such as putting a comb into a bag or placing a mug upside down on a tray, and SO-101 clips, both without success rates. Launch partners run it on a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot.

Fine tuning it yourself

The repository documents full fine-tuning of praxis-1-base for a new rig, with DROID as the worked example, and LoRA adaptation of praxis-1-droid on one 80 GB GPU. LeRobot versions of the checkpoints are on Hugging Face. For the SO-100 and SO-101 there is a ready checkpoint, praxis-1-so101, fine-tuned on about 1,070 hours from 3,814 public community datasets; it predicts five joints and the gripper at 15 Hz in 1.6 s chunks and runs with or without a wrist camera. The RT-1 fine-tune took 30k steps, about 28 hours on 32 H100s.

Where it helps, where it does not

Strengths

  • Best open-weight result on RoboLab-120: 46.2 percent, ahead of every open-weight VLA and world action model on the board.
  • Apache-2.0 for weights and code, commercial use included.
  • Adapts to a robot it has not seen: 80.0 percent on SimplerEnv after fine-tuning on RT-1 data, against 52.4 percent for RT-1 itself.
  • Ready checkpoints for the DROID Franka, bimanual YAM arms and the SO-100 and SO-101, with LeRobot support.
  • Holds up under 4-bit quantization: 46.5 against 47.8 percent on 600 RoboLab-120 episodes.

Limits

  • Hardware: one GPU with at least 80 GB for inference, and about 0.55 to 0.6 s per call on an H100.
  • Reaches for look-alike objects, the most common real-robot failure according to the report.
  • Vague instructions and multi-step tasks from a single prompt remain hard.
  • The real-robot comparison and the SimplerEnv number are Runway's own measurements. Four entries rank above it on RoboLab-120, two frontier model agents and two closed models.
  • No success rates are published for the SO-101 checkpoint.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-11-22.

Try a policy on a real arm

A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.