Kimi K3 is pitched as a near-3-trillion-parameter model that wants a 64-GPU rack
teortaxesTex · x · 2026-07-21
A quoted thread discusses Kimi K3’s scale and deployment footprint, arguing that the model is so large it needs nearly a full rack of high-end chips to run.
From the quoted material
- K3 is described as nearly 3 trillion parameters.
- The poster suggests it needs roughly a full rack, with 64 GPUs rather than necessarily 72.
- The attached technical slide says Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes).
- It also says K3 uses Stable LatentMoE, activating 16 of 896 experts.
- The slide claims quantization-aware training from SFT onward, with MXFP4 weights and MXFP8 activations.
- For inference efficiency, it recommends deployment on supernode configurations with 64 or more accelerators.
- The slide further mentions a corresponding vLLM implementation and says KDA with prefill cache allows competitive token pricing despite the model’s scale and long context.
In short, the post combines a model-scale claim with concrete inference-hardware guidance.
Related event: Moonshot Releases 2.8T Open-Weights Model Kimi K3(14 posts)→
More from Infra
- Nebius secures $3.5B Citi financing facility for GPU purchases and AI data centers — Beth_Kindig · 2026-10-26
- Nearly half of Claude revenue now flows through cloud marketplace commitments, up from 11% — BenBajarin · 2026-10-26
- Anthropic ARR goes vertical after price cuts: tokens up 4x, $100B+ ARR back in play — GavinSBaker · 2026-10-26
- Alex Banks argues power caps GPU supply, real risk is demand cooling first — thealexbanks · 2026-10-26
- Analyst: Ant has the most fungible AI workload of any model provider today — BenBajarin · 2026-10-26
- Anthropic S-1: training infra edge cut per-task cost 40% and doubled revenue growth — BenBajarin · 2026-10-26