SSyncDev

Praxis-1

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Praxis-1 is Runway's open 27B world action model for robotics, predicting video and robot actions together; Apache 2.0, LeRobot and vLLM-Omni support.

Key specifications

Specification score
20.0
License
Apache-2.0

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Weights free under Apache 2.0. Hosted inference via Reactor and a Runway Dev trial; no published prices found.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Image input
  • Video input
  • Text input

Produces

  • Video output

Capabilities

VisionOpen weights

What is Praxis-1?

Runway is known for video generation. Praxis-1 turns that video pretraining toward robots. It is a world action model: a 27B-parameter mixture-of-experts diffusion transformer that predicts future video and robot actions together, conditioned on a natural-language instruction such as "put the mug on the tray". The open release was announced on 11 November 2026, after a first unveiling in September.

The licence is the headline for robotics teams. Weights, inference code and fine-tuning code are all Apache 2.0, commercial use included.

Checkpoints and robots

The release is a collection, not one file:

  • a base checkpoint (about 27.2B parameters in its safetensors);
  • embodiment checkpoints for DROID, SO-101 and YAM setups;
  • LeRobot-format copies of those;
  • NVFP4 variants for smaller hardware.

Repositories are gated with automatic approval. Launch partners report adapting one checkpoint, with light fine-tuning, to a legged humanoid, a six-axis industrial arm and a dual-arm warehouse robot. Full fine-tuning and LoRA are both supported, with NVIDIA NeMo Automodel tutorials for post-training.

Hardware range

Runway lists NVIDIA hardware from Jetson Thor (on the robot) through workstation cards like the RTX 5090 and RTX PRO 6000 to H100 and B200 in the data centre. Expect the NVFP4 builds to be the realistic choice for anything onboard.

Reported results

Benchmark Praxis-1 Comparison
RoboLab-120, 1,200 episodes 46.2% FLUX 3 Action 42.9%, Cosmos3 Nano 36.8%, pi0.5 28.0%
SimplerEnv, after fine-tuning on RT-1 80.0% original RT-1 policy 52.4%

These are Runway's own runs. The RT-1 result is notable because that robot was not in the training data. One outside assessment notes it cannot yet plan complex multi-step action sequences, so long-horizon tasks still need a planner above it.

Integrations

Native support in LeRobot and vLLM-Omni means you do not have to write a serving stack. Hosted inference runs on Reactor, and a trial is available on Runway Dev; we did not find published prices for either.

How it fits next to reasoning models

A planner such as Gemini Robotics-ER 1.6 Preview breaks a task into steps and reasons about the scene. Praxis-1 sits lower: it turns an instruction and camera frames into motor actions and an imagined future. Many stacks will want both.

Getting a first result

Start with the embodiment checkpoint nearest your hardware. Collect fifty to a hundred teleoperated demonstrations of one task in your own space, fine-tune with LoRA and compare success against your current policy over at least fifty trials. Log failure types, not just the rate. Inspect the predicted video when it fails; seeing what the model expected to happen is one of the practical advantages of a world action model. Related categories: reinforcement learning models and image-and-text-to-video models.

Several repositories went up a week before the public release; we use the release date.

How to evaluate Praxis-1

Compare the model on the work you need to complete.

Tasks and inputs

Praxis-1 is listed for Image-text-to-video, Reinforcement learning. It accepts image, video, text and produces video. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Choose duration, resolution and frame rate together. Compare the cost of an accepted clip, including repeat generations.

Choosing a deployment

The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

An open world action model for robotics: a 27B mixture-of-experts diffusion transformer that predicts video and robot actions together from instructions and camera input.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

DocPereira

DocPereira/PEAL_V4_LHP_Zero_Entropy_Controlled

DocPereira/PEAL_V4_LHP_Zero_Entropy_Controlled is a source-linked Hugging Face repository assigned to Reinforcement learning. Its recorded task interface accepts environment state and produces action.

— context

edbeeching

edbeeching/decision-transformer-gym-hopper-medium

edbeeching/decision-transformer-gym-hopper-medium is a source-linked Hugging Face repository assigned to Reinforcement learning. Its recorded task interface accepts environment state and produces action.

— context

HumanCompatibleAI

HumanCompatibleAI/ppo-Pendulum-v1

HumanCompatibleAI/ppo-Pendulum-v1 is a source-linked Hugging Face repository assigned to Reinforcement learning. Its recorded task interface accepts environment state and produces action.

— context

HumanCompatibleAI

HumanCompatibleAI/ppo-seals-CartPole-v0

HumanCompatibleAI/ppo-seals-CartPole-v0 is a source-linked Hugging Face repository assigned to Reinforcement learning. Its recorded task interface accepts environment state and produces action.

— context