Running Kimi K2.7 Code on GB300 NVL72

Kimi K2.7 Code has 1000 billion parameters. GB300 NVL72 offers 20,700 GB of HBM3e across a 72-GPU rack memory. This page puts the two side by side.

By QDNA · Published on 27 July 2026 · Updated on 15 September 2026

Short answer. Kimi K2.7 Code fits comfortably on GB300 NVL72. In FP16, its weights take about 2,300 GB of the 20,700 GB available, leaving 18,400 GB for the attention cache and concurrency.

How much memory does Kimi K2.7 Code need?

Weights take the parameter count multiplied by the format size, plus a 15% runtime margin for activations and buffers. In FP16 the calculation reads: 1,000 billion parameters × 2 bytes (FP16) = 2,000.0 GB of weights; × 1.15 runtime margin = 2,300.0 GB. The attention cache sits on top: it grows with context length and with the number of concurrent requests, so it is sized case by case.

Format	Weights in memory	On this machine	What the format costs
FP16	2,300 GB	fits	full precision, the quality reference
FP8	1,150 GB	fits	negligible loss on most tasks
NVFP4	575 GB	fits	Blackwell format, native FP4 compute
Q4	690 GB	fits	block quantisation, broadly supported

Model sheet: 1,000 billion parameters, 32 billion active per token, 262,144-token context, Modified MIT licence. 1 trillion announced (model card), 8 of 384 experts active; the repository holds 1,027 billion elements, published in 4-bit (595 GB).

Kimi K2.7 Code activates 32 billion parameters per token out of 1000 billion. Active parameters govern speed, not footprint: every expert stays resident in memory. Confusing the two under-sizes the machine by an order of magnitude.

What is left to serve requests?

Kimi K2.7 Code fits comfortably on GB300 NVL72. In FP16, its weights take about 2,300 GB of the 20,700 GB available, leaving 18,400 GB for the attention cache and concurrency. The remaining margin decides how many concurrent requests and how much context the machine sustains. By QDNA convention a machine “fits” when the weights take at most 75% of its memory; beyond that it “barely fits” and forces a limit on context or on concurrency. The attention cache is not calculated on this page. These models’ architectures (compressed latent, linear or sparse attention, Mamba layers) share no per-token formula; it is measured on the machine, with the target context and concurrency [TO BE MEASURED].

GB300 NVL72: price on quotation. NVIDIA publishes no price; analyst estimates range from $3.7M to $6.5M per rack (Loop Capital, Tom’s Hardware, as reported by Spheron on 16 August 2026). The machine targets the GE segment.

Where else can Kimi K2.7 Code run?

See the Kimi K2.7 Code and GB300 NVL72 fact sheets.

How much memory does Kimi K2.7 Code need?

In FP16, weights take about 2,300 GB including the runtime margin. The attention cache sits on top and depends on context.

Does Kimi K2.7 Code fit on GB300 NVL72?

Kimi K2.7 Code fits comfortably on GB300 NVL72. In FP16, its weights take about 2,300 GB of the 20,700 GB available, leaving 18,400 GB for the attention cache and concurrency.

Method and sources

Memory footprints are calculated, not measured: a reading on real hardware may differ depending on the engine and the exact weight format. Every figure below carries its source and the date of its last check.

Kimi K2.7 Code: 1,000 billion parameters, 32 billion active, 262,144-token context, Modified MIT licence. Hugging Face repository (config.json, API, model card), checked on 2 September 2026. 1 trillion announced (model card), 8 of 384 experts active; the repository holds 1,027 billion elements, published in 4-bit (595 GB).
GB300 NVL72: 20,700 GB of memory; 72 Blackwell Ultra GPUs and 36 Grace CPUs, 20 TB of GPU memory per NVIDIA (72 × 288 GB = 20,736 GB), 37 TB fast memory, 130 TB/s NVLink, 1,440 PFLOPS FP4 with sparsity (1,080 dense); power draw not published. Manufacturer sheet checked on 2 September 2026.
GB300 NVL72 price: on quotation. NVIDIA publishes no price; analyst estimates range from $3.7M to $6.5M per rack (Loop Capital, Tom’s Hardware, as reported by Spheron on 16 August 2026).
Bytes per parameter: FP16 2 (16 bits per parameter, hence 2 bytes); FP8 1 (8 bits per parameter (E4M3), hence 1 byte, block scales not counted); NVFP4 0.5 (4 bits per parameter, hence 0.5 byte; 4-bit published weights weigh 0.54 to 0.56 byte per parameter with their scales (DeepSeek V4 Pro 865 GB for 1,599 billion, Kimi K3 1,561 GB for 2,780 billion)); Q4 0.6 (taken here as llama.cpp Q4_K_M, 0.6 byte per parameter including scales; Q4_K_S weighs 0.56, AWQ and GPTQ 0.55). Source: connaissance/faits.yaml, quantifications family, checked on 2026-08-31.
Runtime margin × 1.15: QDNA operating assumption, not measured: 15% above the weights for activations, buffers and fragmentation.
“Fits” threshold at 75% of memory: QDNA assumption, not measured, which keeps the remaining quarter for the attention cache and concurrency.