Runway Opens Praxis-1 Robot Model Weights, at 46.2% on RoboLab-120

Runway released the weights of Praxis-1, a 27-billion-parameter robot model, on Nov. 11 under the Apache 2.0 licence, which allows commercial use. The model reaches 46.2% success on NVIDIA's RoboLab-120 leaderboard, which means it fails more than half of its 1,200 test episodes.

Runway introduced Praxis-1 in September. Its release post says the model is a mixture-of-experts diffusion transformer that predicts video and actions together, and that it builds on Runway's video pretraining. The Robot Report wrote up the release on Nov. 12 and quotes Andy Chen, head of robotics at Runway.

What Runway released

The package holds the weights, inference code and fine-tuning code. Full fine-tuning and LoRA are both supported. Runway lists GitHub, Hugging Face with native LeRobot support, vLLM-Omni, and hosted inference through its Reactor service. The Robot Report said hosted pricing is not published.

Runway names five supported chips: H100 and B200 data-centre GPUs, RTX 5090 and RTX PRO 6000 workstation GPUs, and Jetson Thor for on-robot use.

The RoboLab-120 numbers

Runway's figures, which it attributes to the NVIDIA leaderboard as of November 2026:

1. Praxis-146.2 %
2. FLUX 3 Action42.9 %
3. Cosmos3 Nano36.8 %
4. π0.528 %

Source: Runway release post, Nov. 11, 2026, accessed 2026-11-14

Runway calls Praxis-1 the top-performing world action model on the board and says it beats every open-weight vision-language-action model. The Robot Report wrote that the model ranks fifth overall, so the claim holds within one model category and not across the whole leaderboard.

A second figure comes from SimplerEnv. After fine-tuning on Google's RT-1 robot, which was absent from training data, Praxis-1 reached 80.0% success against 52.4% for the original RT-1 policy. Both numbers are vendor-supplied.

Launch partners and what they report

Runway named three launch partners: Noble Machines, which makes a legged humanoid for heavy industrial handling; Standard Bots, a maker of six-axis factory arms; and Ultra, which builds dual-arm warehouse packing stations.

Chen told The Robot Report that Praxis-1 learned a warehouse plank-picking task from half an hour of examples and succeeded 15 times out of 15. On a messy pile of unseen planks it cleared the whole pile six times out of 10.

Those results come from a partner test that nobody outside has reproduced. Runway's post does not list failure modes or the training data mix beyond general video.

Why web video matters here

Runway's research page reports that web video works about as well as teleoperated robot video for pretraining a policy. If that holds, the data bottleneck that has pushed labs towards expensive teleoperation shrinks. It is one lab's finding and has not been replicated in the sources reviewed.

The context is a crowded field. OpenAI disbanded its robotics team as the site reported, while AgiBot's G2 humanoids have run store trials in Shanghai. Runway said its next phase is scaling deployments with partners at real sites, and has reported no results from that yet.