Unverified50% confidenceFactExact time

Gemma 4采用混合专家架构（MoE），E2B后缀表示Effective 2B，每次推理只激活约20亿有效参数

1

Sources

50%

Confidence

Long-term

Relevance

7/14/2026

First Seen

Sources

GDScript+Vulkan在Godot中运行Gemma大语言模型实验

Related Claims

VerifiedGemma 4的26B MOE模型采用混合专家架构，拥有260亿总参数但任意时刻仅约40亿参数处于激活状态77% similar UnverifiedGemma 4的E2B版本在社区基准测试中多项任务表现堪比Gemma 3的271亿参数模型77% similar UnverifiedGemma 2B模型拥有约20亿参数72% similar UnverifiedGemma 26B-A4B采用Top-K路由策略，每个token只激活K个专家（通常K=2）68% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/504401

API

curl https://kongchang.com/api/v1/knowledge/claims/504401

MCP

get_claim(id=504401)