JD open-sources JoyAI-VL-Interaction, a full-stack interactive vision-language model
京东全栈开源JoyAI-VL-Interaction,从"一问一答"走向"边看边说"
JD released JoyAI-VL-Interaction, the first fully open-source interactive model with native vLLM-Omni support. It watches a live video stream, decides when something matters, and speaks up in real time—complex tasks can be handed off to a backend agent. In a 58-person blind test, it beat Doubao's video call assistant 77.6% of the time and Gemini's 87.9%, hitting 100% on surveillance alerts. The release includes weights, the interaction dataset, training recipes, and a deployable system that takes camera or live-stream input with voice interaction and long-term memory, targeting security, elder care, and live commentary.
Why it matters: JD open-sourced its full JoyAI-VL-Interaction stack; the model decides on its own when to speak in a video stream and beat Doubao and Gemini in a 58-person blind test with 77.6% and 87.9% win rates.