Unverified60% confidenceFactExact time

三元表示理论上每个参数只需约1.58比特，相比FP16可以带来近10倍的存储压缩

2

Sources

60%

Confidence

Long-term

Relevance

8/5/2026

First Seen

Sources

Maple-Preview：20B三元MoE模型iPhone端侧推理达120tok/s

Related Claims

Verified4-bit量化将每个参数压缩至0.5字节，理论上将140GB压缩至约35-40GB78% similar Unverified量化可将700亿参数模型的存储体积从FP16的约140GB压缩至INT4的约35-40GB，能在单张RTX 4090配合CPU内存卸载的消费级硬件上运行77% similar UnverifiedParquet列式存储压缩比通常达行式格式的3-10倍75% similar Verified1750亿参数模型仅模型权重（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1TB75% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/692441

API

curl https://kongchang.com/api/v1/knowledge/claims/692441

MCP

get_claim(id=692441)