Runway Open-Sources Praxis-1 27B Robot World-Action Model

Runway has officially launched Praxis-1, a 27-billion-parameter mixture-of-experts diffusion transformer designed for robot control, releasing open weights and code under an Apache 2.0 license that permits commercial use. Officially released on November 12, 2026, the model represents Runway's first entry into robotics AI and delivers state-of-the-art performance on industry benchmarks while remaining accessible to researchers and commercial developers alike.

Architecture and Design

Praxis-1 combines a 26.9-billion-parameter video backbone with a 16-layer action expert, totaling 27.2 billion parameters with approximately 2.8 billion active per forward pass. The architecture is a mixture-of-experts diffusion transformer that predicts a short video sequence of the robot's next-step future before deciding on motion commands. Each inference call generates camera video alongside 24 action rows spanning 1.6 seconds at 15 Hz, providing temporal context for decision-making.

The model ships in bfloat16 format at 54.4 GB, fitting on a single 80 GB GPU. Runway also released NVFP4-quantized variants whose routed experts run in 4-bit precision on SM120 architecture GPUs, reducing memory requirements while maintaining inference quality. On Hugging Face, the company provides base, DROID, and bimanual YAM variants to support different robotic platforms and manipulation tasks.

Benchmark Performance

On NVIDIA's RoboLab-120 leaderboard, Praxis-1 achieves 46.2% success across 1,200 test episodes, outperforming competing models by significant margins:

• FLUX 3 Action: 42.9% (3.3 points behind)
• Cosmos3 Nano: 36.8% (9.4 points behind)
• Physical Intelligence π0.5: 28.0% (18.2 points behind)

In a cross-platform generalization test, Runway fine-tuned Praxis-1 on Google's RT-1 robot dataset, which was not included in the model's original training data. On the SimplerEnv benchmark, the fine-tuned version reaches 80.0% success compared to 52.4% for the original RT-1 policy, a 27.6 percentage point improvement that demonstrates strong transfer learning capabilities across robotic platforms.

Training Data and Methodology

The training mix for Praxis-1 comprises 58% egocentric human video, 20% general video, and 18% real robot video. Runway reports that web video performs comparably to teleoperated robot video as pretraining data for policy performance, a finding that could reduce data collection costs for future robot learning systems. Fine-tuning used approximately 277 hours of successful manipulation episodes from the DROID dataset, captured with Franka robotic arms.

This approach follows recent research showing that large-scale video pretraining, even on non-robotic data, can bootstrap visual understanding and physical intuition that transfers to robot control tasks. The model learns temporal dynamics and object interactions from passive video before specializing on robot trajectories.

Release and Availability

Runway released the complete Praxis-1 package under Apache 2.0 licensing, explicitly permitting commercial deployment without royalties or usage restrictions. The Hugging Face repository includes model weights, inference code, and quantized variants optimized for different hardware configurations. This release follows Runway's Solaris model in September 2026 and marks the company's first contribution to open-source robotics research.

The published benchmark scores represent Runway's internal evaluations measured against other vendors' reported figures rather than third-party replications. Standard practice in AI research involves such comparisons, though independent verification typically follows in subsequent academic studies.

What This Means

Praxis-1's release lowers the barrier to entry for robotics AI development by providing a commercially viable, high-performance foundation model. The combination of strong benchmark results, flexible licensing, and hardware-efficient quantized variants positions the model as a practical starting point for robotic manipulation research and product development. The 27.6-point improvement over RT-1's original policy on transfer tasks suggests that large-scale video pretraining can accelerate robot learning across platforms, potentially reducing the need for extensive per-robot data collection in future systems.