Open-weight robotics and world models

Runway opens Praxis-1 robot-action weights and Apache-2.0 code

A publicly downloadable video-action policy now comes with base, DROID and YAM checkpoints, but its robotics success rates remain publisher-reported.

Open Model Weights published12 Nov 2026
Primary sourceRunway — official Praxis-1 public-weights release
Source published2026-11-11
FOLLOW OUR COVERAGE

Follow Open Model Weights on Google See more of our open-weight AI reporting in Google Search.

Add as a Preferred Source on Google
Original widescreen futuristic engineering illustration of an industrial robotic arm and a sequence of predicted video-action frames, with clear Praxis-1 editorial headline.
Original conceptual editorial illustration; not real test footage, a hardware photograph or an independent benchmark. AI-generated editorial illustration; not a documentary photograph.

Runway has publicly released its Praxis-1 world action model, following an initial announcement and robotics-partner trials in September. The November 11 release adds downloadable weights for a base checkpoint, a single-arm policy fine-tuned on DROID and a bimanual YAM policy. The company also published inference, simulation and full- and LoRA-fine-tuning code under Apache-2.0, including commercial use. The newsworthy change is inspectable, adaptable model artifacts, not merely another robot demonstration or a promise that weights will appear later. Primary source ↗

Praxis-1 predicts video and robot actions jointly. Its technical report describes a roughly 27.2-billion-parameter system: a 26.9B video mixture-of-experts backbone, around 2.8B of whose parameters are active per token, plus a smaller action expert that reads the video model's hidden states. In the documented single-arm setting, a policy call receives camera observations, a textual instruction and robot state, then produces a future-video chunk together with action commands. Subsequent chunks contain 24 steps at 15 Hz, equivalent to 1.6 seconds. The system separately uses a frozen Wan2.2 video autoencoder and Qwen3.5-2B text encoder.

Runway reports 46.2% task-episode success on NVIDIA's RoboLab-120 simulation benchmark, spanning 1,200 trials, compared with its cited 42.9% for FLUX 3 Action. The company also reports adaptation to Google's RT-1 hardware and an 80.0% SimplerEnv success figure. Those are publisher-reported results for particular policies, fine-tunes and protocols. They do not mean Praxis-1 ranks above every robotics agent: Runway's report lists higher-scoring frontier-model-driven agents that use different approaches. Open Model Weights has not reproduced these benchmarks or measured real-world robot reliability.

The released BF16 backbone and action checkpoint are about 54.4 GB, and Runway documents an 80 GB GPU for baseline deployment. NVFP4 variants and embedded platforms introduce their own hardware assumptions. Teams considering an actual robot should first check the precise checkpoint SHA, weight access, license texts and third-party dependency terms, then test action interfaces and physical safety constraints on their own embodiment. A News announcement does not add an unverified checkpoint to our model registry or confer a Leading classification.

OMW REGISTRY WATCH

What this changes for the evidence layer

New high-relevance robotics source candidate; no model registry admission. Verify runwayml/praxis-1-base, praxis-1-droid and praxis-1-yam SHAs, accessible artifacts, component licenses and fine-tuning context. All RoboLab-120 and SimplerEnv numbers remain publisher claims until independently reproduced.

PRIMARY SOURCE

Runway — official Praxis-1 public-weights release

This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.

Open source article ↗