The promise from September is now a download
When Runway introduced Praxis-1 in September, the open release was a commitment, not a product. Three launch partners were running it on their own machines, and everyone else had a blog post. That changed on November 11. The weights are on Hugging Face, the code is on GitHub, and the license is Apache 2.0, which allows commercial use. The release covers the inference and fine-tuning code as well, with both full fine-tuning and LoRA — adapting the model without retraining all of it — supported.
This is the part of the announcement that holds up on its own terms. A video company has taken the video pretraining behind its general world models, built a model that outputs robot actions alongside predicted video, and handed it to anyone with the hardware to run it. In a field where several recent "open" video releases arrived with licenses that exclude whole countries or reserve commercial use, an uncomplicated Apache 2.0 is itself the news.
First place, at 46.2% of episodes
Runway calls Praxis-1 the top-performing world action model on NVIDIA's RoboLab-120 leaderboard, and on the numbers it published, that is accurate. Across 1,200 episodes it succeeded 46.2% of the time. FLUX 3 Action managed 42.9%, Cosmos3 Nano 36.8%, and Physical Intelligence's π0.5 28.0%.
Read the winning number on its own terms. In Runway's reported evaluation, 53.8% of the 1,200 benchmark episodes were unsuccessful — for the highest-scoring model in the comparison. That is not an argument against the release. It is a measure of where robot policy models actually are on this benchmark, stated plainly by the company in first place. Anyone quoting "state of the art" from this announcement without the 46.2% next to it is quoting half the result.
Two further cautions belong with the score. The figures are Runway's, marked "data as of November 2026," and the sources reviewed provide no independent replication of them. And the comparison set is the models on that one leaderboard, in the configurations Runway lists. The download changes who can check this: because the weights and code are public, labs outside the partner program can now rerun the benchmark instead of citing it.
The more useful number is 80%
The result that says more about the model's purpose is not the leaderboard. Fine-tuned on Google's RT-1, a robot Runway says never appeared in Praxis-1's training data, the model reached 80.0% success on the SimplerEnv simulation benchmark, against 52.4% for the original RT-1 policy. Runway's argument is that large-scale video pretraining transfers: a model that has watched enough of how objects move can be adapted to drive a body it was not built for. With its launch partners, Runway says, a single checkpoint has adapted to a variety of embodiments with light fine-tuning.
That claim also comes from Runway, and SimplerEnv is a simulator, not a warehouse floor; the result concerns simulated evaluation, not measured factory reliability. But it is the claim the open release actually tests. If outside teams can reproduce the transfer on their own robots, Praxis-1 becomes a starting checkpoint for physical AI rather than a Runway product. Independent evaluation should distinguish reproduction of the published results from transfer to new robots and tasks; the two are separate questions, and the release now lets outsiders answer both.
What to watch
Useful next evidence includes independent RoboLab-120 runs, now possible for any lab with the hardware; adaptation reports from teams outside the three launch partners, especially on hardware below the H100 tier Runway leads with; and whether the partner robots move from demonstrations to paid work. Until then, the license is the settled part of this release: it permits inspection, modification and commercial use of the released materials. It does not establish performance or production readiness — those are what the downloads are for.
Sources
- [1] Runway, "Praxis-1 Is Now Available" (official announcement, November 11, 2026)Read source
- [2] Runway, Praxis-1 open-source repository (GitHub, Apache-2.0)Read source
- [3] Runway, Praxis-1 base weights (Hugging Face)Read source
- [4] Daily AI Brief, "Runway Releases Praxis-1 Robot Model With Open Weights" (November 11, 2026)Read source
- [5] ALFW, earlier coverage of the September introduction (context, different event)Read source