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.

Original post →

Kimi K3 is pitched as a near-3-trillion-parameter model that wants a 64-GPU rack

More from Infra