# Kimi-K3 GPU requirements

Kimi-K3 has 2.8T parameters. To serve it, plan for about 6,672 GB of GPU memory in 16-bit, 3,336 GB in 8-bit or 1,668 GB in 4-bit, including about 20% for the KV cache and activations.

It is a mixture of experts: 896 experts, 16 of them active for each token. All of them stay in GPU memory, so plan memory for all 2.8T parameters.

[Deploy it on Arrakis](/console/#/models?model=moonshotai/Kimi-K3).

## What it takes to serve it

| Precision | Memory and the smallest setup |
|---|---|
| 16-bit | 6,672 GB: more than one 8-GPU machine |
| 8-bit | 3,336 GB: more than one 8-GPU machine |
| 4-bit | 1,668 GB: 8 × [B300](/gpus/b300.html.md), from $36.80 per hour for all 8 |

## What it takes to fine-tune it

| Method | Memory and the smallest setup |
|---|---|
| QLoRA | 1,946 GB: 8 × [B300](/gpus/b300.html.md), from $36.80 per hour for all 8 |
| LoRA | 6,672 GB: more than one 8-GPU machine |
| Full fine-tuning | 44,479 GB: more than one 8-GPU machine |

## About the model

| | |
|---|---|
| Parameters | `2.8T` |
| Experts | `896` |
| Experts for each token | `16` |
| Context window (tokens) | `1,048,576` |
| Published in | `FP8` |
| License | `other` |
| Released | `2026-06-13` |
| Weights | [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3) |

## Questions

**How much GPU memory does Kimi-K3 need?** To serve it, about 6,672 GB in 16-bit, 3,336 GB in 8-bit, 1,668 GB in 4-bit, including about 20% for the KV cache and activations. To fine-tune it, about 1,946 GB with QLoRA, 6,672 GB with LoRA and 44,479 GB for full fine-tuning with Adam, before activations.

**Can I run Kimi-K3 on one GPU?** No. Even in 4-bit it needs about 1,668 GB: 8 B300 GPUs.

**What does it cost to run Kimi-K3?** In 16-bit it needs more than one 8-GPU machine; in 4-bit, 8 B300 GPUs. Each setup on this page shows the lowest price per GPU-hour on Arrakis right now, and you pay by the hour from a prepaid balance.

**How do I deploy Kimi-K3 on Arrakis?** Open Models in the Arrakis console, paste moonshotai/Kimi-K3, choose vLLM, SGLang or llama.cpp, and rent a GPU it fits on. You get an OpenAI-compatible endpoint with its own API key, and root SSH to the machine.


More models: [/models](/models.html.md).