# Kimi-K2.5 GPU requirements

Kimi-K2.5 has 1.0T parameters. To serve it, plan for about 2,465 GB of GPU memory in 16-bit, 1,232 GB in 8-bit or 616 GB in 4-bit, including about 20% for the KV cache and activations.

It is a mixture of experts: 384 experts, 8 of them active for each token. All of them stay in GPU memory, so plan memory for all 1.0T parameters.

[Deploy it on Arrakis](/console/#/models?model=moonshotai/Kimi-K2.5).

## What it takes to serve it

| Precision | Memory and the smallest setup |
|---|---|
| 16-bit | 2,465 GB: more than one 8-GPU machine |
| 8-bit | 1,232 GB: 8 × [B200](/gpus/b200.html.md), from $54.32 per hour for all 8 |
| 4-bit | 616 GB: 8 × [H100](/gpus/h100.html.md), from $13.20 per hour for all 8 |

## What it takes to fine-tune it

| Method | Memory and the smallest setup |
|---|---|
| QLoRA | 719 GB: 4 × [B200](/gpus/b200.html.md), from $27.16 per hour for all 4 |
| LoRA | 2,465 GB: more than one 8-GPU machine |
| Full fine-tuning | 16,430 GB: more than one 8-GPU machine |

## About the model

| | |
|---|---|
| Parameters | `1.0T` |
| Experts | `384` |
| Experts for each token | `8` |
| Context window (tokens) | `262,144` |
| License | `other` |
| Released | `2026-01-01` |
| Weights | [moonshotai/Kimi-K2.5](https://huggingface.co/moonshotai/Kimi-K2.5) |

## Questions

**How much GPU memory does Kimi-K2.5 need?** To serve it, about 2,465 GB in 16-bit, 1,232 GB in 8-bit, 616 GB in 4-bit, including about 20% for the KV cache and activations. To fine-tune it, about 719 GB with QLoRA, 2,465 GB with LoRA and 16,430 GB for full fine-tuning with Adam, before activations.

**Can I run Kimi-K2.5 on one GPU?** No. Even in 4-bit it needs about 616 GB: 8 H100 GPUs.

**What does it cost to run Kimi-K2.5?** In 16-bit it needs more than one 8-GPU machine; in 4-bit, 8 H100 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-K2.5 on Arrakis?** Open Models in the Arrakis console, paste moonshotai/Kimi-K2.5, 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).