Kimi K3's compute load too heavy to serve economically, analysis argues
teortaxesTex · x · 2026-11-04
Analysis by YouJiacheng (amplified by teortaxesTex) on Kimi K3's serving economics:
- K3's large KV cache doesn't block serving at compute-bound batch sizes — the bottleneck isn't memory bound but raw compute volume
- Even counting only 6x active parameters without CED, Kimi can't serve it economically
- Dense MLA's compute burden makes it worse: at 77k context it's equivalent to 400B active parameters
- Conclusion: this is a compute/cost problem, not an infra-efficiency problem
More from Infra
- SpecAgent: multi-session Claude Code pipeline with measured 3x context cost savings — AdOtherwise9458 · 2026-11-05
- SoftBank cloud arm IDC Frontier hit by ransomware in 'seven minutes,' taking ~500 organizations offline in Japan — ChuckDBrooks · 2026-11-05
- A $2.5k local rig runs Qwen3.8 27B at 150 tokens/s, benchmarks show — casper_hansen_ · 2026-11-05
- A $2.5k RTX 3090 box runs 328GB GLM-5.3-Flash locally via expert offloading — casper_hansen_ · 2026-11-05
- Arm earnings: data center royalties double as selling own chips pits it against licensees — tengyanAI · 2026-11-05
- Open-source Local Office adds Snapdragon NPU support for local AI on Windows ARM64 — HowDevelop · 2026-11-05