What GPU do I need to run moonshotai/Kimi-K2-Instruct?
A 1026B (MoE) language model for chat and instruction-following. 1026.4B parameters, published in F8_E4M3. View on Hugging Face
Kimi-K2-Instruct is published by moonshotai on Hugging Face, with 166,449 downloads and 2,376 likes to date. It's a DeepseekV3ForCausalLM model built for text-generation, published natively in F8_E4M3.
What Kimi-K2-Instruct is
Kimi-K2-Instruct is a 1026B-parameter mixture-of-experts language model published by Moonshot AI on Hugging Face. It is released under Custom license.
License note: a lab-specific license (tagged "other" on Hugging Face); read the model card's own license section before commercial use. Facts in this section are sourced from Kimi-K2-Instruct's Hugging Face model card, not benchmarked by Aquanode.
What it's used for
- Chat assistants
- Instruction following
- Synthetic data generation
VRAM required & cheapest live GPU fit
Required VRAM = weight size at each precision times a flat 1.2 overhead for activations and fragmentation. The KV-cache grows with context and is not in that factor; it is listed per model below where the architecture is published. Full formula and assumptions: methodology.
| Precision | Weight size | Required VRAM | Cheapest live fit | GPUs needed | Est. $/hr (full fit) |
|---|---|---|---|---|---|
| FP8 (native) | 955.9 GB | 1147.1 GB | AMD MI300X | 6 | $15.77/hr |
| INT4 (quantized) | 478.0 GB | 573.6 GB | AMD MI300X | 3 | $7.89/hr |
A GPU is only matched to a row if its hardware supports that precision, and the primary recommendation is always a single-GPU fit when one exists.
INT4 requires a quantized checkpoint actually published for this model, check its Hugging Face page before relying on this row.
Cheapest way to run Kimi-K2-Instruct at its published (F8_E4M3) precision: 6× AMD MI300X, at $2.63/hr per GPU ($15.77/hr total). Quantizing to FP8 or INT4 (rows above) can cost less, but requires a compatible quantized checkpoint to exist for this model.
Kimi-K2-Instruct: KV cache by context length
The KV cache is the memory the attention layers hold for every token of context, on top of the weights. It is not part of the flat 1.2x overhead in the table above, grows with context length and with every concurrent request, and is why a long-context deployment needs more VRAM than the table shows.
Multi-head latent attention: every layer caches one compressed latent per token. Cached per token: 70,272 bytes at 16-bit.
| Context | KV cache, one sequence |
|---|---|
| 4K | 0.27 GB |
| 32K | 2.1 GB |
| 128K | 8.6 GB |
Computed from the layer, head and window counts in the model's own configuration (moonshotai/Kimi-K2-Instruct), at 16 bits per cached value. Each concurrent sequence needs its own cache, on top of the weights.
Kimi-K2-Instruct: common questions
Can Kimi-K2-Instruct run on a single GPU?
No. At FP8 (native) it needs 1147.1 GB of VRAM, more than a single 32 GB desktop card holds. The cheapest capable card in the live feed is a 192.0 GB AMD MI300X, and it takes 6 of them.
Is Kimi-K2-Instruct already quantized?
Yes. It is published in FP8, one byte per parameter, so the 1147.1 GB figure is already a quantized footprint rather than a full-precision one. Only INT4 goes below it, at 573.6 GB. A GPU without FP8 tensor cores cannot run it as published, which is why cards here are matched on precision support and not on VRAM alone.
How many GPUs do I need to run Kimi-K2-Instruct?
6 at FP8 (native). It needs 1147.1 GB of VRAM and the cheapest capable live offer is a 192.0 GB AMD MI300X, so 6 of them come to $15.77/hr in total.
Does quantizing Kimi-K2-Instruct lower the GPU bill?
Yes. At FP8 (native) the cheapest live fit is 6 AMD MI300X cards at $15.77/hr. At INT4 (quantized) it drops to 3 AMD MI300X cards at $7.89/hr, provided a quantized checkpoint exists for it.
How to run Kimi-K2-Instruct
Run Kimi-K2-Instruct with vLLM
From moonshotai/Kimi-K2-Instruct's own deployment docs.
# start ray on node 0 and node 1
# node 0:
vllm serve $MODEL_PATH \
--port 8000 \
--served-model-name kimi-k2 \
--trust-remote-code \
--tensor-parallel-size 16 \
--enable-auto-tool-choice \
--tool-call-parser kimi_k2Source: https://huggingface.co/moonshotai/Kimi-K2-Instruct/raw/main/docs/deploy_guidance.md
Run Kimi-K2-Instruct with GGUF quantizations
Prebuilt GGUF weights published at unsloth/Kimi-K2-Instruct-GGUF. Run with llama.cpp's llama-server or load the repo directly in LM Studio.
llama-server -hf unsloth/Kimi-K2-Instruct-GGUFSource: https://huggingface.co/unsloth/Kimi-K2-Instruct-GGUF
Deploy Kimi-K2-Instruct on Aquanode
Aquanode has no one-click deploy template for Kimi-K2-Instruct; you install the inference engine yourself with the commands below. Aquanode sells GPU pods billed per second, not a hosted inference API.
- Launch a bare GPU pod sized to the requirement above (6× AMD MI300X or larger).
- Open a terminal on the pod, or save one of the commands above as a startup script so it runs automatically the first time the pod boots.
- Run the command and connect to the resulting endpoint.
Weight-to-VRAM math, the fit rules, and how live prices are normalized: full methodology.
More Kimi K2 models
- Kimi-K2-Base (1026.5B, F8_E4M3)
- Kimi-K2-Instruct-0905 (1026.5B, F8_E4M3)
- Kimi-K2.5 (1026.9B, Native INT4)
- kimi-k2.6-eagle3-mla (3.0B, BF16)
Alternatives at this size
Other models for text-generation within about a third of Kimi-K2-Instruct's 1026.4B parameters, from other model lines.
- GLM-5.2 (753.3B, BF16)
- A.X-K2 (691.7B, F8_E4M3)
- MiMo-V2.5-Pro (1023.2B, F8_E4M3)
- Hy4-preview (780.0B, BF16)
- DeepSeek-V4.1-Flash-1017-obliterated (754.6B, BF16)
More on Kimi-K2-Instruct
Best chat and assistants models: how Kimi-K2-Instruct ranks against the rest.
Compare Kimi-K2-Instruct with DeepSeek-R1, DeepSeek-V3.2, DeepSeek-V3-0324, DeepSeek-V3.
Related reading: AMD MI300X pricing and specs, How much VRAM you need for LLMs, Serving LLMs with vLLM, vLLM vs TensorRT-LLM vs SGLang, and Best GPU for LLM inference.
Your own GPU cloud accounts
One console for the GPU clouds you already rent from. Your keys stay on your machine.
$20 a month, flat