DeepSeek’s New AI Is A Game Changer
Two Minute Papers1580 字 (约 7 分钟)
87
DeepSeek’s visual pointing lets open-source VLMs slash visual tokens by 90 % while matching or beating GPT-4V on seven public benchmarks and delivering traceable reasoning paths.
入选理由：视觉指针机制将视觉 token 用量压缩 90%，仍保持 SOTA 精度
FeaturedVideo#DeepSeek#Vision-Language Models#Visual Pointing#Token Efficiency#Open Research英文