Unverified50% confidenceFactExact time
对于1.6万亿总参数、490亿激活参数配置，实际推理时只有约3%的参数参与计算
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGrok 1 的3140亿总参数中每次推理仅激活约25%，实际推理算力接近650亿参数规模的密集模型78% similarUnverified1000亿参数的模型处理单次推理请求需约80GB显存，百亿参数级别模型仅需约20GB77% similarUnverified通义千问3.8约有两万亿参数，Kimi约有两万亿参数77% similarUnverifiedKimi 2.5每次前向传播只计算约300亿参数，尽管总参数量达1万亿71% similarVerified若MoE模型每次推理仅激活约1/8的参数，实际计算开销可能仅相当于百亿参数级别的稠密模型71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/912187API
curl https://kongchang.com/api/v1/knowledge/claims/912187MCP
get_claim(id=912187)