Unverified50% confidenceBenchmarkExact time
改变块大小时MMLU多选题几乎不受影响仍稳在78%，说明知识本身还在，崩掉的是生成过程
1
Sources
50%
Confidence
Long-term
Relevance
7/20/2026
First Seen
Sources
Related Claims
UnverifiedPerformance degradation caused by quantization is most pronounced in mathematical reasoning, code generation, and precise factual recall tasks, while conversation, summarization, and creative writing are relatively robust.64% similarVerifiedEven models with a million-token context face an attention dilution problem when handling large codebases—when the context is too long, the model's attention to early content drops significantly.62% similarUnverified响应速度快无法证明背后是否为官方正版服务，能力越强的模型每次推理消耗算力越多、响应时间相应更长61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/565269API
curl https://kongchang.com/api/v1/knowledge/claims/565269MCP
get_claim(id=565269)