Unverified50% confidenceFactExact time
Embedding模型的向量维度范围从768维到3072维不等，影响精度与存储成本的平衡
1
Sources
50%
Confidence
Long-term
Relevance
9/6/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedEmbedding模型输出的高维向量通常为768至3072维，向量数据库通过HNSW等图算法可将亿级向量检索压缩至毫秒级响应76% similarUnverified雅可比透镜的内存需求随模型规模线性增长，大致为模型体积的1/866% similarVerified4-bit量化理论上可将模型体积缩小8倍（相比32位存储）65% similarUnverified7MB模型在语义表达细腻度上很难与数百MB的重量级模型相比，激进压缩伴随精度损失64% similarUnverified1750亿参数模型仅模型权重（BF16精度）就需350GB存储，Adam优化器需维护一阶矩和二阶矩两份状态，合计超过1TB63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/864030API
curl https://kongchang.com/api/v1/knowledge/claims/864030MCP
get_claim(id=864030)