Runway Releases Praxis-1 Robot Model With Apache 2.0 Weights Behind a Hugging Face Access Gate; NVIDIA's RoboLab-120 Board Lists It Fifth of 21

Overview

Runway has released Praxis-1, a robot-control model it calls a world action model, with weights and inference and fine-tuning code under the Apache 2.0 licence, The Robot Report reports. Runway’s launch post is dated November 11, 2026. The Robot Report describes the model as a 27B-parameter mixture-of-experts diffusion transformer that predicts video and actions together. On NVIDIA’s RoboLab-120 simulation leaderboard it scores 46.2%, which both The Robot Report and The Decoder place fifth overall; The Decoder also puts it at the top of the world action model category. The Hugging Face model repositories sit behind an access prompt, and The Decoder says the results have not been independently verified while the leaderboard’s own page for Praxis-1 carries a “Verified” badge.

What was released

According to The Robot Report, Praxis-1 is available on GitHub, on Hugging Face with native LeRobot support and in vLLM-Omni, with hosted inference on Reactor. Runway said the release covers the weights and the inference and fine-tuning code, with support for full fine-tuning and low-rank adaptation (LoRA), and that the Apache 2.0 terms apply “including for commercial use”.

The model pages add a qualification. The Hugging Face pages for praxis-1-base and praxis-1-droid carry the tag “License: apache-2.0” and a licence section reading “Apache License 2.0. The frozen dependencies (Wan2.2 VAE, Qwen/Qwen3.5-2B) are distributed separately under their own licenses.” The GitHub README (its latest commit, per the GitHub API, is dated November 12, 2026 at 22:14 UTC) says the code is under Apache 2.0 and adds: “See the license file of each set of weights for the weights.”

Both Hugging Face pages (base, droid), as the Herald read them on November 14, show the notice “This repository is publicly accessible, but you have to accept the conditions to access its files and content.” and a line saying users must agree to share their Hugging Face information. The pages show the conditions themselves only after logging in, so the Herald has not read them. A Herald query of the Hugging Face API on November 14 listed ten Praxis-1 model repositories under the runwayml organisation; each was flagged as gated and tagged Apache 2.0. The Decoder reports that developers can download the model “after signing up and agreeing to the access terms”.

The base checkpoint is a starting point, not a ready policy. Its card says it has no per-embodiment action heads, “so it does not run as a policy as is”, and The Decoder says the base model needs additional training before it works with new robot systems. According to the README, the released single-arm checkpoint is for a DROID rig and the bimanual one for two YAM arms, with NVFP4 versions and an SO-101 checkpoint also listed.

Hardware requirements

The Robot Report writes that Praxis-1 runs out of the box across NVIDIA hardware, from H100 and B200 data-centre GPUs to the RTX 5090, RTX PRO 6000 and Jetson Thor. The README quick start needs one CUDA GPU with at least 80 GB of memory, and the praxis-1-droid card says “Inference needs one 80 GB GPU.” For the optional NVFP4 path the README says the option “has not been run in closed loop on an SM120 GPU yet” (it lists the RTX 5090 as an SM120 GPU) and that a 32 GB RTX 5090 “should fit (the peak is not measured yet)”. These statements describe different configurations (the default bf16 weights and the optional NVFP4 path), so the Herald does not treat them as contradicting each other.

Results and who ran them

The Robot Report relays Runway’s RoboLab-120 figures: 46.2% success across 1,200 episodes, against 42.9% for FLUX 3 Action, 36.8% for Cosmos3 Nano and 28.0% for π0.5. The leaderboard row shows 555 of 1,200 episodes at 10 per task. Runway’s launch post says the model “outperforms every open-weight VLA”.

The Herald’s reading of the leaderboard’s “All” view on November 14 listed 21 policies and showed no last-updated time. Four sit above Praxis-1: AMP1 (86.0%) and SenseNova-RoboRSI (57.1%), both typed “Agentic LM”; Lakeside (51.0%), typed LWM; and BiMind v0.5 (49.2%), typed VLA. The board marks Lakeside and BiMind v0.5 with “Closed source: model weights not publicly available”. The Decoder describes it as “behind two AI agents and two robot models with different architectures”. Among rows typed WAM or VLM+WAM, Praxis-1 has the highest success rate; FLUX 3 Action follows at 42.9%.

Andy Chen, the head of robotics at Runway, told The Robot Report that on a Noble Machines warehouse task Praxis-1, in Runway’s own test, “learned it from half an hour of examples and got it right 15 times out of 15”.

Per The Decoder, Runway’s research report adds two more comparisons. On a real Franka arm, both Praxis-1 and FLUX 3 Action attempted ten tasks ten times each, and Praxis-1 succeeded 88 times to FLUX 3 Action’s 80. After fine-tuning on about 95 hours of robot data from the mobile robot Google used for RT-1, an unseen platform, Praxis-1 averaged 80 percent in simulated tests against 52.4 percent for Google’s RT-1 model.

Verified or not

The Decoder writes: “Runway ran all the tests itself, and the results haven’t been independently verified yet.” The leaderboard’s detail page for Praxis-1 shows a “Verified” badge, a submission date of November 5, 2026 and the tooltip “RoboLab organizers were able to independently run this policy and/or cross-check its reported results”. The “and/or” leaves open which of the two was done. The Decoder’s sentence appears in a report that also covers real-robot and RT-1 tests, which the board does not. No cited source explains the difference, so the Herald reports both. The leaderboard is NVIDIA’s, per The Decoder; The Robot Report lists NVIDIA hardware as what Praxis-1 runs on, and Runway’s launch post says “Runway and NVIDIA are opening frontier technology stacks”.

What we don’t know

• The Decoder says whether the approach needs less robot data than comparable models without general video pretraining “is still an open question”.
• None of the cited sources reports an outside replication of the real-robot or RT-1 comparisons, which The Decoder attributes to Runway’s own tests.
• The peak memory on an RTX 5090 is not measured, per the README.
• The Decoder, citing Runway’s research report, says the model sometimes confuses similar-looking objects, struggles with vague instructions and cannot reliably plan complex action sequences on its own.