JD.Com Takes Its AI Ambitions Physical With Two New JoyAI Models at WAIC

JD.com showed up to WAIC 2026 in Shanghai this week with a clear message: AI is about to leave the screen.

The Chinese e-commerce and tech giant, better known for logistics and retail than language models, unveiled two new AI models at the conference on July 18 under its JoyAI brand. JoyAI-Talker handles real-time voice interaction with emotional awareness, while JoyAI-Video-Edit lets users edit video footage by describing what they want in plain language.

JoyAI-Talker is the more interesting of the two. JD says it can sustain low-latency conversation, pick up on emotional cues in a user’s voice, call external tools, and even maintain memory across interactions. It identifies users through voiceprint recognition and factors in environmental data to understand conversational context: the difference between “turn it up” in a quiet office versus a noisy warehouse matters when the AI is supposed to actually help.

The voice model represents a shift from AI that mechanically follows commands to one that tries to understand intent and state. It’s a crowded space: Baidu, Alibaba, and Tencent all have competing voice assistants, but JD’s edge might be in application: the company runs massive logistics operations, customer service centers, and retail warehouses where voice interaction with context awareness could be genuinely useful.

JoyAI-Video-Edit takes a different swing. It treats video as a live, editable stream rather than a finished file. Users look at a changing video frame and tell the model to add or remove objects, swap people, change clothing, or restructure an entire scene, all through natural language and no timeline editing required. The technology is flashy, though practical utility for anything beyond simple edits remains to be seen.

These launches cap a busy year for JD’s AI efforts. The company has now released seven base models in 2025-2026, including Joy-Image-Edit, JoyAI-Echo, and JoyAI-VL-Interaction, a real-time visual-language interaction model that JD open-sourced earlier.

At WAIC, JD also released EgoLive, described as the industry’s largest human-perspective dataset, collected via head-mounted cameras. Conference attendees could try out a wearable JoyEgoCam and experience first-person data collection at JD’s “JoyInside AI Home” exhibit, a physical demonstration space that wraps all these models into a single environment.

JD’s strategy is starting to look less like a retailer experimenting with AI and more like a company building a full-stack AI platform spanning voice, image, video, real-time interaction, world models, and embodied intelligence. Whether that translates into products people actually want to use will depend on how well these models work outside the controlled environment of a conference hall.