YouJiacheng 详算 K3 与 V4.1F 的 KV cache:每 token 14400 字节对 890 字节
YouJiacheng · x · 2026-10-26
Responding to z4y5f3's perf modeling claims, YouJiacheng provides the concrete numbers: K3 has 24 MLA layers, each using 600 bytes/token with FP8 KV, totaling 14400 bytes/token, versus V4.1F's 890 bytes/token — revising the claimed 40x KV cache gap down to 16x. Read together with the parent post in this dispute over serving costs.
所属事件:Kimi K3 推理成本之争：解码算力悬殊，贵 DeepSeek V4.1F 约 20 倍(6 条相关)→
「Infra」频道最新
- Anthropic S-1：4300 亿美元算力合同不可取消，超过中国全年军费 — Hesamation · 2026-10-27
- Tenstorrent 融资 15 亿美元，估值 62 亿较去年翻倍以上 — Beth_Kindig · 2026-10-27
- Anthropic 拟向 SpaceX 采购 1205 亿美元算力,满载后月付约 31 亿 — XFreeze · 2026-10-27
- 恶搞：用 JEV 决策模型替换 Raft 共识，写入吞吐暴跌 99.98% — irombie · 2026-10-27
- Cadence 进入领域专用 AI 模型赛道,芯片设计自动化再加码 — karlfreund · 2026-10-27
- Qdrant 自研 Universal I/O:向量库 mmap 无缝迁移到对象存储 — qdrant_engine · 2026-10-27