Unverified60% confidenceFactTime unknown
Gemma 4 26B MOE版本在24GB显存GPU上实际推理速度达到每秒60个Token
2
Sources
60%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026
Sources
Related Claims
UnverifiedGemma 4 26B MOE版本可在24GB显存的显卡上运行78% similarUnverifiedGemma 4 31B 经量化后模型大小约为 20GB，实测 GPU 显存占用约 20.5GB，GPU 利用率约 91%78% similarUnverifiedGemma 3 12B的FP16版本约需24GB显存，Q4_K_M版本降至约7-8GB，推理速度提升约2-3倍，结构化输出准确率仅损失约2-5%77% similarUnverifiedGemma 4 31B can run on a regular consumer GPU with around 50 TOPS of compute.73% similarVerifiedGemma 4的1B/2B模型量化后仅需4GB显存即可运行73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/57145API
curl https://kongchang.com/api/v1/knowledge/claims/57145MCP
get_claim(id=57145)