How JD's JoyAI Model Matrix Is Driving AI Into the Physical World

JD argues that the next AI competition shifts from generating text, images, and video to understanding space, remembering tasks, and coordinating devices in the physical world. Rather than chasing model size or traffic, JD leverages two decades of supply‑chain data and real‑world scenarios to continuously train, validate, and apply AI in industry and daily life.

During the 2026 World Artificial Intelligence Conference (WAIC), JD showcased the JoyAI model matrix, including the real‑time speech interaction model JoyAI‑Talker (low‑latency dialogue, emotion understanding, tool calling, memory) and the real‑time video editing model JoyAI‑Video‑Edit (on‑the‑fly editing and preview). These models extend JD’s existing suite—Joy‑Image‑Edit, JoyAI‑Echo, JoyAI‑VL‑Interaction—forming a foundation that covers voice, image, video, interactive world modeling, and embodied intelligence.

To address the bottleneck of high‑quality embodied data, JD mobilized 600,000 participants to collect 10 million hours of first‑person video, creating the industry‑largest human‑viewpoint dataset EgoLive , which it open‑sourced. The data pipeline turns raw video into annotated “textbooks” for robots, builds digital twins of shelves, and enables robots such as the Xingchen intelligent sorting robot to perform real‑world tasks like shelf stocking.

JD also introduced JoyInside “AI Home” , a household ecosystem where devices—from smart lamps that project and answer questions to AI‑controlled tea makers and sleep‑monitoring mattresses—share a common AI brain that can identify speakers, understand intent, retain long‑term memory, and coordinate actions across multiple appliances. Over 200 brands have joined, and more than 10 million devices are slated for integration this year.

The overall architecture links data collection, simulation training, and on‑device verification, providing a full‑stack infrastructure covering “collect, store, label, train, evaluate, simulate, test.” JD’s cloud AI platform supplies the engineering backbone for model training, deployment, and large‑scale application across logistics, health, finance, industrial supply chains, and smart homes.

By open‑sourcing both models and embodied data, JD aims to lower the entry barrier for the industry, enabling other companies to move AI out of the lab and into real‑world operations.