There is a substantial hardware catch, however. Runway advertises RTX 5090 support, but its repository distinguishes the standard installation from an experimental quantized route. “Runs locally” and “validated on your desktop GPU” are not interchangeable claims.
From predicting video to choosing actions
Runway describes Praxis-1 as a 27-billion-parameter mixture-of-experts diffusion transformer. Its premise is that large-scale video pretraining can teach useful patterns about objects and task sequences before adaptation to particular robots. The company initially introduced the model in September; November’s announcement opens access more broadly.
The implementation makes the idea more concrete. A policy call receives a task instruction, camera observations and the robot’s current pose, then predicts a video segment alongside a corresponding action segment. The robot executes that action chunk. This is a chunked robot-control policy, rather than evidence of unrestricted understanding of the physical world.
The RTX 5090 qualification matters
The repository’s default quick start requires Python 3.12, a CUDA 13.0 driver and a CUDA GPU with at least 80GB of memory. It provides single-arm DROID and dual-arm YAM checkpoints, plus simulation and fine-tuning workflows. Those are developer prerequisites, not the requirements of an ordinary desktop application.
For compatible SM120 GPUs, including the RTX 5090, the repository documents an optional NVFP4 inference path. Crucially, it warns that quantization changes action outputs and that this path had not yet been run closed-loop on an SM120 GPU. That caveat deserves equal billing with the consumer-GPU headline: loading a model is a different milestone from demonstrating reliable robot behavior with it.
For PC developers evaluating the release, the practical implication is straightforward: choose the execution path first, then assess its memory requirements and validation status. Do not assume results reported for one hardware and precision configuration automatically transfer to another.
What the benchmark establishes—and what it does not
Runway reports 46.2% success across 1,200 RoboLab-120 episodes, against 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5. These are company-reported comparisons, not independent confirmation of universal superiority.
The benchmark’s research paper supplies essential context: RoboLab-120 is a simulation benchmark covering 120 tasks, organized around visual, procedural and relational competencies across three difficulty levels. Its authors designed it to investigate generalization and sensitivity to controlled changes.
Consequently, 46.2% should not be read as the percentage of arbitrary household jobs Praxis-1 can perform. Nor does a simulation leaderboard establish safe operation around people. The useful finding is narrower: under the reported evaluation conditions, the model leads the comparison Runway presents.
A developer release, not a Windows robot assistant
Praxis-1’s immediate value is inspectability and adaptation: developers receive weights, code and documented testing workflows rather than only a hosted demonstration. The repository’s simulation route offers a sensible starting point before physical hardware enters the equation. It does not establish a supported, turnkey Windows installation.
The release is therefore meaningful for robotics and GPU-software teams—but the fine print is part of the product story. Open weights lower an access barrier. They do not remove the engineering between an impressive prediction and a robot that consistently does the right thing.