Tag
UltraViT is a latency-optimized vision encoder for large vision-language models, designed for on-device deployment with a pyramidal architecture and a two-stage generative pre-training strategy, achieving state-of-the-art performance at 1.7x speed.