Unverified50% confidenceBenchmarkExact time
260 亿参数的 Gemma（MoE）实测速度接近 31 Token/秒，比 GPT-OSS 20B 快约三分之一
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/19/2026
First Seen
Valid until: 12/18/2026
Sources
Related Entities
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65% similarUnverifiedMuse Glimmer 30B在RTX 5090上报告可达每秒74.9个token，评测者在RTX 3090上运行全精度版本观察到约60-65 token/秒65% similarVerifiedM4芯片的Neural Engine峰值算力达到38 TOPS（每秒38万亿次运算）63% similarUnverified经过精心量化的4-bit模型与FP16原始模型之间的性能差距通常在1-3个百分点以内63% similarUnverifiedGemma 4 26B MOE版本总参数量为25.2B，但推理时仅激活约3.8B参数63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/933721API
curl https://kongchang.com/api/v1/knowledge/claims/933721MCP
get_claim(id=933721)