UniVR-34B-Planning is a 34B multimodal model built on Emu3.5 that performs unified next-token prediction directly in visual space. It learns complex reasoning, physical dynamics, and multi-step planning from image-based demonstrations without intermediate text chains or task-specific heuristics. Released openly on Hugging Face with checkpoints, code, and dataset, it targets agentic and robotic workflows such as manipulation and control.

Read the official ByteDance launch page

First builds incoming — this lane fills in automatically as people start building with UniVR-34B-Planning. Know one? Submit the URL.