How much VRAM does Kimi K2 Instruct need?

About 650 GB at Q4_K_M with a 32K-token context (628 GB weights, 2.3 GB KV cache, 19.8 GB overhead). These are estimates.

Can I run Kimi K2 Instruct on a single GPU?

Not at Q4_K_M with 32K context on any single GPU in our catalog. Use several GPUs, a high-memory unified-memory machine, a smaller quantization or a shorter context.

Which apps can run Kimi K2 Instruct locally?

Verified so far: llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang. Each entry links to its source on this page.

Models / Kimi K2 Instruct

Can I run Kimi K2 Instruct locally?

Moonshot AI · 1.03T parameters · mixture of experts, 32B active per token · up to 128K context · license other · Hugging Face

At Q4_K_M with a 32K context, Kimi K2 Instruct needs about 650 GB of GPU or unified memory by our estimate. The official checkpoint (FP8) is 1029 GB on disk.

Memory by quantization and context

Quantization	Weights	Total at 8K	Total at 32K	Total at 128K
Q3_K_M (~4.0 bit)	513 GB	530 GB	531 GB	538 GB
Q4_K_M (~4.9 bit, popular default)	628 GB	648 GB	650 GB	657 GB
Q5_K_M (~5.7 bit)	732 GB	755 GB	757 GB	764 GB
Q6_K (~6.6 bit)	842 GB	869 GB	871 GB	878 GB
Q8_0 (~8.5 bit, near-lossless)	1091 GB	1125 GB	1127 GB	1134 GB
F16 / BF16 (16 bit)	2053 GB	2116 GB	2118 GB	2125 GB

Bits per weight from llama.cpp's quantize docs. F16 KV cache. Totals include our runtime-overhead estimate.

Cheapest hardware that fits (Q4_K_M, 32K)

NVIDIA DGX Station (GB300): about €83,738 (new, configurator.exxactcorp.com, 740 GB usable)

Runs in

Tool	Platforms	How	Source
llama.cpp Inference engine (C/C++)	Linux, Windows, macOS, Android (build)	GGUF: unsloth/Kimi-K2-Instruct-GGUF	source
MLX (mlx-lm) Apple framework + CLI	macOS (Apple Silicon only)	MLX conversion: mlx-community/Kimi-K2-Instruct-4bit	source
KTransformers CPU+GPU hybrid engine	Linux, Windows	named in the official model card	source
vLLM Serving engine	Linux	named in the official model card	source
SGLang Serving engine	Linux	named in the official model card	source

Only tools with a public source are listed: an official build (ggml-org, lmstudio-community, mlx-community, bartowski or Unsloth on Hugging Face), the model card, ollama.com or the tool's own README. Apps built on llama.cpp, such as Jan and PocketPal AI, can usually load the same GGUF files. All local apps.

Try other settings

ModelQuantizationContext lengthKV cacheMax GPUs

Estimated memory needed650 GB

Weights 628 GBKV cache at 32K 2.3 GBRuntime overhead 19.8 GB

1 hardware options fit

Cheapest real offer per device from AI Hardware Deals, converted to EUR at approximate rates. Multi-GPU rows multiply the card price.

Also show hardware that doesn't fit Include legacy server cards (M40, P40, P100, MI50, MI60)

	Setup		Headroom
NVIDIA DGX Station (GB300) unified memory	1 device	740 GB	89.9 GB	€83,738 new, configurator.exxactcorp.com	≤ 363 tok/s

Speed ceiling is memory bandwidth divided by the bytes of (active) weights read per token, shown only where we have a sourced bandwidth figure. Real decode speed is lower, especially with long contexts or when splitting across cards. All memory numbers are estimates; see the methodology.

FAQ

How much VRAM does Kimi K2 Instruct need?: About 650 GB at Q4_K_M with a 32K-token context (628 GB weights, 2.3 GB KV cache, 19.8 GB overhead). These are estimates.
Can I run Kimi K2 Instruct on a single GPU?: Not at Q4_K_M with 32K context on any single GPU in our catalog. Use several GPUs, a high-memory unified-memory machine, a smaller quantization or a shorter context.
Which apps can run Kimi K2 Instruct locally?: Verified so far: llama.cpp, MLX (mlx-lm), KTransformers, vLLM, SGLang. Each entry links to its source on this page.