Verne releases open weights for V0.5, a robot model built for factory and warehouse work

Robotics startup Verne released V0.5 on November 4, 2026, an open-weight "world action model" aimed at industrial robots. The weights are on Hugging Face behind an access request, under Verne's own V0.5 Model License. Verne asks commercial users to contact it for a partnership.

V0.5 has 8.2 billion parameters and starts from the Wan2.2-TI2V-5B video model. Verne pretrained it on more than 1.7 million demonstrations from data vendors, open-source datasets, its own teleoperation and its handheld grippers. In training, a video transformer learns how a scene changes while an action transformer learns the motion. At runtime only the action side runs, outputting chunks of 32 targets at 30 Hz for the arms, grippers and mobile base.

Verne published three checkpoints. The base model is a starting point for fine-tuning, not for direct deployment, and the release has no training or fine-tuning code; Verne points users to its own fine-tuning platform. Verne says its fine-tuned versions reach 99.3% average success on LIBERO and 81.0% clean / 73.4% randomized on RoboTwin 2.0, and calls V0.5 the leading open-weight model from a US company on both, as of November 1, 2026. On LIBERO's average, one other model, ABot-M0.5, scores slightly higher.

Why it matters

Verne says it already deploys robots in semiconductor cleanrooms, cold storage and large warehouses, and that a larger next model is training. With open weights, other teams can start from a pretrained model instead of training one from scratch, though they still need demonstrations from their own robot to fine-tune it. The benchmark results come from simulation, and Verne compares itself against published numbers for models such as Physical Intelligence's π0.5.

Sources: Verne on X, Hugging Face model card, Verne: Introducing V0.5.