Documentation

Everything you need to get started with CloudGPU.

rocket_launchQuick Start Guide

1

Create an Account

Sign up at CloudGPU with email or Google — no card needed.

Create an Account →
2

Go to Deploy

Open the Deploy page and pick a model — DeepSeek, Qwen, Llama, or Mistral.

Go to Deploy →
3

Click Deploy

Select a GPU, click Deploy. We handle everything — GPU setup, model download, API server.

4

Copy the API Endpoint

You get an OpenAI-compatible API URL ready to use.

5

Paste Into Your Code

Set base_url to your endpoint. That's it — no per-token charges, you pay for GPU time.

Paste Into Your Code →

smart_toyLet your AI coding assistant rent the GPU

The CloudGPU MCP server lets Cursor, Cline, opencode, Claude Code and other MCP clients quote a price, rent a GPU once you approve, run your job, download the results and shut the GPU down. CloudGPU's servers enforce the budget and time limit on every instance.

Set up the MCP server →

apiUsing your instance as an API endpoint

For developers building AI agents, chatbots, or any app that needs an LLM backend with a fixed hourly cost instead of per-token billing.

Every running Ollama instance gives you an OpenAI-compatible endpoint. Open the instance in your Dashboard and click the API button to copy-paste one of these:

Replace <random-id> with the id shown on your instance row.

OpenAI Python SDK

from openai import OpenAI client = OpenAI( base_url="https://<random-id>-i.cloudgpu.app/v1", api_key="ollama", # any non-empty string ) resp = client.chat.completions.create( model="qwen2.5:0.5b", messages=[{"role": "user", "content": "Hello"}], ) print(resp.choices[0].message.content)

curl

curl https://<random-id>-i.cloudgpu.app/api/chat \ -H "Content-Type: application/json" \ -d '{ "model": "qwen2.5:0.5b", "messages": [{ "role": "user", "content": "Hello" }], "stream": false, "keep_alive": "30m" }'

Tip: keep_alive

Pass "keep_alive": "30m" in every request to keep the model in VRAM for 30 minutes after the last call. Without it, Ollama unloads the model after 5 minutes of idle and your next request pays a 30–90 second cold-load penalty.

memoryWhich GPU should I pick?

All prices below are per-hour retail.

Model sizeMin VRAMRecommended GPUPrice
0.5B – 8B (Qwen 2.5, Llama-3.1-8B, Mistral-7B)16 GBRTX 4090 24 GB$0.39/hr
13B – 32B quantized (Qwen-32B-AWQ, CodeLlama-34B)32 GBRTX 5090 32 GB$0.69/hr
30B+ at 8-bit, or 70B at 4-bit with short context48 GBRTX 4090 48 GB$0.69/hr
32B at 4-bit with long context, or 13B–32B at 8-bit with short context40 GBA100 40G$0.79/hr
70B at 4-bit with long context (Llama 3.3 70B AWQ, Qwen2.5-72B-Instruct-AWQ), or 8-bit with short context80 GBA800 80G$1.19/hr
Large-context LLM + long reasoning96 GBH20 96G$1.29/hr
Enterprise inference / multi-agent production80 GBH800 80G$1.79/hr

Some cards can be rented several to one instance (up to 8): pick the number of cards on the Deploy page, and the hourly price multiplies by the count.

helpFrequently Asked Questions

How does billing work?expand_more
You are billed per minute of GPU time, with a 1-minute minimum. While an instance runs, each started hour is pre-charged from your balance; when you destroy it, the unused minutes go back to your balance. All prices are in USD. Billing rules and examples
What's the minimum deposit?expand_more
The suggested minimum top-up is $5. There is no signup credit.
Can I stop and restart my instance?expand_more
You can destroy an instance at any time. Currently, stopped instances release their GPU — you'll need to create a new instance to resume. Persistent storage across restarts is on the roadmap. What to save before you destroy an instance
What is pre-installed on an instance?expand_more
Every template is a ready-made image with root SSH: Ollama (empty, or with Qwen, DeepSeek-R1, Llama, Gemma, Mistral, GLM or Phi preloaded), JupyterLab + Ollama, ComfyUI (download FLUX or other weights yourself), or LLaMA-Factory. Templates that serve an API (Ollama, ComfyUI) also get a public HTTPS URL. The ComfyUI template installs ComfyUI only: it does not include FLUX or other model weights, so download the weights you need after the instance is ready.
Is there an API?expand_more
Yes. Every instance exposes an OpenAI-compatible REST API (for LLM templates) or model-specific REST API (for image / audio templates). See the API button on any running instance in your dashboard for copy-pasteable curl, Python and Node examples.
How do I get support?expand_more
Email support@cloudgpu.app or visit our Contact page. We aim to respond within 24 hours while the service is in beta.
What's the difference between Templates and Deploy?expand_more
Templates is the catalogue — search and filter everything you can run, and see what is ready versus still coming. Deploy is where you actually launch: pick a template or a bare GPU, pick a card, and you get a URL. Browse in Templates, launch in Deploy.
Can I deploy custom models?expand_more
For LLMs, our Ollama template supports any model in the Ollama library — after deploy, use the 'Pick Model' button to pull anything from qwen2.5:0.5b up to llama3.1:70b. For anything else, rent a raw GPU (Deploy → Raw GPU) and install what you need over SSH.
How do I connect my deployed model to my application?expand_more
Open your running instance in the dashboard and click the API button. You'll get an OpenAI-compatible endpoint at /v1/chat/completions plus ready-to-paste code for curl, Python (requests + OpenAI SDK), and Node.js. Drop it into LangChain / AutoGen / CrewAI with one line.
Downloads from Hugging Face or GitHub are slow. What can I do?expand_more
Run cloudgpu-net on in the instance's terminal. It turns on download acceleration for that shell, for Hugging Face, GitHub, Civitai and similar sites. pip and other downloads can be slower while it is on: run cloudgpu-net off to turn it off, and cloudgpu-net status to check. In a startup script, run cloudgpu-net on before your downloads. If the machine has no acceleration, the command says so.

Still have questions?

Contact Support