DeepSeek VL2

DeepSeek open MoE vision-language model — efficient, strong OCR and doc understanding.