Meta ships Muse Spark 1.2: image/video-to-working-code plus robotic tool-use loops

AI at Meta detailed Muse Spark 1.2 across multimodal coding and robotics. The model can turn images or video into working web pages and games, scoring itself on actual rendering and behavior in a self-improvement loop. A specialized variant acts as a robot orchestrator: it parses an instruction, issues tool calls, observes results, and loops until the task is done, including bimanual desk-tidying. Meta also uses it internally with Muse Image/Video for agentic media generation and downstream insight extraction.