Kimi K3 serving costs 20-50x more than DeepSeek V4.1 Flash, perf modeling argues
z4y5f3 · x · 2026-10-26
A technical dispute over serving economics of Kimi K3 vs DeepSeek V4.1 Flash:
- z4y5f3 ran perf modeling at 128K context and found V4.1F uses 40x less KV cache than K3 (not 16x); on B200 EP64 decode, DeepSeek easily hits its compute bound while K3 is one OOM away from saturating compute. Overall K3 is 40-50x more expensive to serve for only 5-10% better quality — a hard value proposition.
- YouJiacheng pushed back: with FP8 KV, K3's 24 MLA layers use 14400 bytes/token vs V4.1F's 890, a 16x gap — but per DeepInfer's published 77K average context, both can reach compute-bound batch sizes on EP64. Still, K3's activations are 6x larger: serving the same traffic would need 21,600 GB200 GPUs vs DeepSeek's 2,400.
- z4y5f3 accepted the FP8 correction (he had used HF default BF16/FP32) but revised the conclusion to K3 still 20x more expensive to serve, noting FP8 doesn't change the compute-vs-bandwidth equation.