Unverified50% confidenceFactExact time
Whisper 的 tiny 或 base 变体参数量从39M到74M不等，配合 INT8 量化可将模型体积压缩至原来的1/4
1
Sources
50%
Confidence
Long-term
Relevance
9/9/2026
First Seen
Sources
Related Entities
Related Claims
Unverified0.9B量级的模型经INT8量化后，理论推理内存占用可降至约1GB以下71% similarUnverifiedint8量化将参数从16位压缩至8位，理论上减少约50%显存占用；int4将显存占用降低至原始的25%左右70% similarUnverifiedINT8量化通过将32位浮点权重映射到8位整数，在几乎不损失精度的前提下将模型体积压缩75%70% similarUnverifiedTiny-MoE模型总参数量约200M，但每个token激活的参数仅约90M68% similarUnverified相比标准INT8量化，三值量化可将模型体积进一步缩小约4倍，并将矩阵乘法退化为纯加减运算66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/881361API
curl https://kongchang.com/api/v1/knowledge/claims/881361MCP
get_claim(id=881361)