Unverified50% confidenceFactExact time
L and M levels barely survive at about 10% coverage; H level can recover from about 30% coverage.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
Unverified传统 LLM 推理中因需按最大长度预留显存，KV Cache 的实际利用率通常不到 30%60% similarUnverifiedThe VRAM usage of a VLM is typically 20%–40% higher than that of a text-only LLM of the same parameter scale.57% similarUnverifiedThe +2 in the denominator of the HN ranking formula is a smoothing term that prevents new posts from receiving abnormally high scores due to an overly small denominator.57% similarUnverified推理时的KV缓存、激活值缓冲区和中间计算状态会额外占用30%-50%显存57% similarUnverifiedA sparse subset comprising only 28%–48% of upstream components is sufficient to retain 88%–101% of the full steering effect.56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1005770API
curl https://kongchang.com/api/v1/knowledge/claims/1005770MCP
get_claim(id=1005770)