Jepa as a Tool

An LLM agent (Muse Spark) that imagines its moves before acting. Before each push, it calls `simulate`, which runs candidate plans through a world model built on frozen V-JEPA 2 and returns the outcomes as text. One world model serves three worlds (2D PushT, 3D tabletop, 3D billiards) and learned from images only. With the world model, the agent's success rate on the 3D tabletop rose from 67% to 89%. It ships with an MCP server, a model on Hugging Face and a live demo in the browser.