1T multimodal MoE for high‑capacity video and image understanding with efficient inference.