Unverified50% confidenceFactExact time
CLIP ViT-L参数量约3亿，嵌入维度为768
1
Sources
50%
Confidence
Long-term
Relevance
7/10/2026
First Seen
Sources
DINOv2 vs SigLIP：k-NN分类为何差距悬殊？视觉编码器选型避坑指南
redditr/MachineLearning7/8/2026
Related Claims
UnverifiedViT-L模型约3亿参数，checkpoint达1213MB，CPU上处理每张图像需4.5秒72% similarUnverifiedViT-S版本嵌入维度为384，为每个16×16图像块输出一个384维特征向量71% similarUnverifiedDINOv2 Giant参数量约11亿、嵌入维度1536，均显著大于SigLIP2 SO400M（约4亿参数、维度1152）和CLIP ViT-L（约3亿参数、维度768），但k-NN任务表现最差68% similarVerified参数量达1750亿的GPT-4级模型，即便以fp8精度存储也需要约175GB显存63% similarUnverified量化后70亿参数模型的显存需求从约28GB压缩至约4-6GB，普通消费级显卡如RTX 3060/4060即可流畅运行Llama 3、Qwen2.5、Mistral等70亿参数级别模型60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/469427API
curl https://kongchang.com/api/v1/knowledge/claims/469427MCP
get_claim(id=469427)