JD opensources JoyAI model stack: assistive AI glasses, interactive world model EchoWM
新智元 · wechat · 2026-09-10
At its JDD conference, JD released and open-sourced a JoyAI foundation model matrix aimed at moving AI from screens into the physical world.
AI glasses and interaction model
- JoyAI-VL-Interaction: an 8B vision-centric interaction model, billed as the first fully open-source interaction model system, fusing perception, reasoning, and interaction decisions in one model to decide in real time when to speak on a live video stream, with sub-second latency; complex tasks are delegated to a background agent.
- In 58 human blind evaluations it beat Doubao's video-call assistant 77.6% and Gemini 87.9% of the time.
World model
- JoyAI-EchoWM: a real-time interactive world model emphasizing interaction fidelity — environments, ambient sound, and speech evolve with user control signals. It scored 81.7 overall / 87.2 interaction on WBench navigation.
Data and embodiment pipeline
- EgoLive: 2,000 hours of 60fps stereo first-person video, 65,866 clips, 346 real-world tasks — claimed as the largest open task-oriented egocentric dataset.
- JoyAI-Sim achieves 90% sim-to-real consistency at 3.2x efficiency; JoyAI-RA0.5 is a VLWA framework using Dual Action Alignment for cross-embodiment manipulation learning.
Industry models
- Jingyi Qianxun 3.0 tops both HealthBench leaderboards with 3D radiology accuracy 15 points above GPT-5.5; JoyIndustrial 2.0 has handled 1B+ industrial calls; VisCAD 1.0 leads general models by 15% on core design tasks.
Strategy
Dr. Duan Nan framed AI's fifth wave as entering the physical world via digital, embodied-on-device, and robotic intelligence, targeting world-class interactive models by end-2026, scaled embodied datasets in 2027, and broad deployment by 2028.
More from Embodied
- Reservoir Farms to showcase crop robots in Salinas on Sept 25 — Scobleizer · 2026-09-10
- Robot vet with 2 billion picks live-streams real parcel induction at 2,200-2,400/hour, dwarfing Figure's 1,200/hour humanoid demo — BradPorter_ · 2026-09-10
- Apple A20 Pro Neural Engine projected to hit 140+ TFLOPS, a third of an A100 for local inference — AIFlow_ML · 2026-09-10
- Dad Clock follows Mom Clock as AI gadget clocks become a family — LuoSays · 2026-09-10
- Six years later, a blogger unboxes a brand-new-in-box Microsoft Surface Duo — revodavid · 2026-09-10
- Puffin-World: unified world model with native 3D states, open-sourced with 16M dataset — ccloy · 2026-09-10