[KongchangAI]
Unverified50% confidenceBenchmarkExact time

Gemma 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%

1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/13/2026
First Seen
Valid until: 10/11/2026

Sources

本地Gemma 3 12B变身AI Agent实战指南：工具调用与框架选择
redditr/aiagents7/12/2026

Related Claims

UnverifiedGemma 4 26B模型约17GB，适合24GB以上显存的显卡，编程效果最佳78% similarUnverifiedGemma 4 26B MOE版本在24GB显存GPU上实际推理速度达到每秒60个Token77% 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%77% similarVerifiedGemma 4的1B/2B模型量化后仅需4GB显存即可运行76% similar

Cite This Claim

Stable URI
https://kongchang.com/claim/501965
API
curl https://kongchang.com/api/v1/knowledge/claims/501965
MCP
get_claim(id=501965)