Verne is releasing V0.5, an open-weight world action model for teams that deploy general-purpose robots in factories, warehouses, and other industrial sites. The model is pretrained to predict both how a scene changes and how a robot should move. At runtime it outputs 30 Hz action chunks in end-effector space for arm poses, gripper commands, and base position. A separate controller turns those targets into motor commands at 500 Hz. A vision-language model sits above V0.5 and breaks longer jobs into short language instructions that the model follows one at a time. The design targets mobile dual-arm work, with the base and both arms moving together so the robot can reach tasks spread across a warehouse floor. Verne pretrained V0.5 on more than 1.7 million demonstrations, starting from the Wan2.2 video model and adapting it for robot control. On evaluations published with the release, V0.5 reaches 99.3% average success on LIBERO. On RoboTwin 2.0 it reaches 81.0% in clean scenes and 73.4% in randomized scenes with clutter, new textures, different lighting, and different table heights. Verne reports that V0.5 was the leading open-weight model from a US company on both benchmarks as of November 1, 2026. The nearest US open-weight comparisons in that post are NVIDIA Cosmos Policy at 98.5% on LIBERO and Physical Intelligence π0.5 at 70.7% clean and 46.0% randomized on RoboTwin 2.0. Founded in 2025 by Neil Nie and Aditya Jha, San Francisco-based Verne (Y Combinator S25) already deploys robots for kitting, assembly, and inspection in semiconductor cleanrooms, cold-storage facilities, and large warehouses. This release lets other teams start from the open weights and adapt them on a new fine-tuning platform using their own demonstrations, rather than training a foundation model from scratch. Verne says its stack can learn a new task, or a new step of a longer task, from under two hours of demonstrations and human corrections.