JoyAI-VL-Preview
JoyAI-VL-Preview (also associated with JoyAI-VL-Interaction-Preview) is an open vision-driven real-time interaction model developed by Joy Future Academy (JD.com Open Source team). Unlike conventional turn-based language models that only respond when directly addressed, this model is engineered to continuously observe live video streams and autonomously decide when to speak, remain silent, or delegate complex tasks.
Built upon an 8-billion parameter architecture derived from robust vision-language foundations (such as Qwen3-VL), the model processes streaming visual input alongside pluggable speech modalities (ASR and TTS). It uses a time-aligned training recipe combined with reinforcement learning to master per-second decision-making. Furthermore, a predictive video codec mechanism helps optimize token consumption over long continuous observation sessions by allocating visual fidelity dynamically.
Create with Crafiq
Generate images, 3D models, video and audio in one studio.
Explore the studioHow JoyAI-VL-Preview ranks
JoyAI-VL-Preview is highlighted in the table below. Switch the metric to see how the ordering changes.