Unverified50% confidenceFactExact time
Gemma 26B-A4B adopts a Top-K routing strategy, activating only K experts per token (typically K=2).
1
Sources
50%
Confidence
Long-term
Relevance
8/16/2026
First Seen
Sources
Related Claims
UnverifiedGemma 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68% similarUnverifiedGemma 4 31B ranks in the top 27 on the overall capability leaderboard, ahead of models including Gemma 2.5 Pro.64% similarVerifiedGemma 4's 26B MoE model uses a Mixture of Experts architecture, with 26 billion total parameters but only approximately 4 billion parameters active at any given moment.61% similarUnverifiedGemma 4's E2B version performs comparably to Gemma 3's 27.1-billion-parameter model on multiple tasks in community benchmark tests.60% similarVerifiedThe Gemma 4 series includes four models: 1B, 2B, 4B, 26B (MOE), and 31B (Dense).59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/756807API
curl https://kongchang.com/api/v1/knowledge/claims/756807MCP
get_claim(id=756807)