Unverified50% confidenceFactExact time
一个32K上下文窗口在Q4量化模型上可能额外占用2-4GB内存
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
Mac本地AI编程代理实战：Ollama+OpenCode+MCP完整搭建指南
redditr/ollama7/11/2026
Related Claims
VerifiedQ5_K_M 量化方案每个权重平均约 5.3 比特，32B 参数模型量化后大约需要 20-24GB 存储空间81% similarVerified4-bit量化（如Q4_K_M）通常可将3B模型内存占用从约6GB压缩至约2GB，精度损失低于2-3%80% similarVerifiedQLoRA将基础模型权重以4bit NF4格式量化存储，使单卡微调130亿参数模型成为可能78% similarVerifiedQ4_K_M 采用混合精度，部分关键权重以 6bit 存储，其余以 4bit 存储，整体平均约 4.5bit77% similarUnverified一张1024×1024的未压缩RGBA纹理占据4MB存储空间75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/490102API
curl https://kongchang.com/api/v1/knowledge/claims/490102MCP
get_claim(id=490102)