Unverified50% confidenceFactExact time
Qwen 2.5 7B使用4组GQA配置，推理时KV Cache体积相比MHA减少约75%
1
Sources
50%
Confidence
Long-term
Relevance
7/14/2026
First Seen
Sources
Related Claims
UnverifiedQwen3.6 27B Q4量化版总分650分，排名第四71% similarVerifiedQwen 2.5拥有包括1.5B、7B、72B在内的多个参数版本71% similarUnverifiedQwen3.5的参数激活率为4.3%，而DeepSeek V3的671B模型激活率为5.5%70% similarUnverifiedGGUF格式的Q4_K_M是最主流的平衡选择，困惑度损失通常低于1%；Q8_0几乎无精度损失但文件体积是Q4的两倍70% similarUnverifiedQwen 3.6 35B A3B在长上下文记忆基准测试中相比Gemma 4和Qwen 3.5具有明显优势68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/506358API
curl https://kongchang.com/api/v1/knowledge/claims/506358MCP
get_claim(id=506358)