Qwen4-Max-Preview
A multimodal reasoning preview for coding, visual analysis and interactive agent workflows through Alibaba Cloud Token Plan.
An open-weight robot control model that combines video prediction and action generation for adaptable robotics policies.
Visit website
Runway Praxis-1 is a world action model for robotics developers and researchers. It applies video pretraining to robot control, taking camera observations, robot state and a task instruction to predict actions.
The release includes model weights, inference code and fine-tuning code under Apache 2.0, including commercial use. Developers can run and adapt the policy on their own hardware, with full fine-tuning and LoRA available for customization.
A 27B-parameter mixture-of-experts diffusion transformer predicts video and robot actions together.
The repository documents DROID single-arm, YAM bimanual and SO-100/SO-101 variants, with checkpoint-specific observation requirements.
Full fine-tuning supports broader adaptation, while LoRA provides a documented route for task adaptation.
The release supplies a policy server and MuJoCo simulation for developing and evaluating robot policies.
Distribution includes GitHub and Hugging Face, with native LeRobot support and vLLM-Omni integration.
Compare task behavior in simulation before planning a physical robot evaluation.
Select an appropriate checkpoint and investigate adaptation to the rig's observations and controls.
Explore LoRA adaptation for a narrowly defined manipulation workflow.
Weights, inference code and fine-tuning code are available under Apache 2.0, including commercial use. Self-hosting requires separately supplied compute and robot hardware; the open release does not include those resources.
Check current pricing ↗Yes. Runway releases the weights and inference and fine-tuning code under Apache 2.0, including commercial use.
Yes. The repository's quick start runs a single-arm checkpoint in simulation, subject to its compute and software requirements.
The documented single-arm policy returns robot commands. Predicted video for the corresponding time span is available when video decoding is enabled; serving configurations that skip decoding still return robot commands.