Open models such as Llama, Qwen, DeepSeek, Mistral, Gemma and OpenAI's gpt-oss can run on one graphics card — if the card has enough memory. Renting one by the hour is the cheapest way to try them: no hardware to buy, and you stop paying the moment you stop.
The short version
- Open One-click AI and choose Chat with an open model. A GPU with enough memory is picked, started and opened in your browser.
- Pick a model from the list (or type any Ollama tag, such as
qwen3:8b), and start chatting. - When you are done, press Stop. The time you did not use comes back to you.
If you would rather choose the machine yourself, use the marketplace: filter by memory (VRAM), price and region, rent the card, and pick the Ollama environment at checkout.
Which GPU do you need?
What decides it is the card's memory (VRAM), not its speed. A model has to fit in memory, with a little room left for the conversation. Quantized versions (4-bit, "Q4") need about a quarter of the memory of the full-precision ones and lose surprisingly little quality.
| Model size | Example models | 4-bit needs | Fits on |
|---|---|---|---|
| 1–4 B | Llama 3.2 3B, Qwen3 4B, Phi | 2–4 GB | any card here |
| 7–9 B | Llama 3.1 8B, Qwen3 8B, Mistral 7B | 5–7 GB | RTX 3060 12 GB and up |
| 12–14 B | Qwen3 14B, Phi-4, Gemma 12B | 9–11 GB | 12–16 GB cards |
| 20 B | gpt-oss 20B | 14–16 GB | 16 GB cards and up |
| 27–34 B | Qwen3 32B, Gemma 27B, DeepSeek-R1 32B | 18–22 GB | RTX 3090 / 4090 (24 GB) |
| 70 B | Llama 3.3 70B, DeepSeek-R1 70B | 40–45 GB | 48 GB cards (RTX 6000 Ada, A6000) or an 80 GB A100/H100 |
| 120 B+ | gpt-oss 120B, bigger MoE models | 65 GB+ | 80 GB cards, or several cards rented together |
An RTX 4090 (24 GB) is the sweet spot for most people: it runs every model up to about 32B quickly, and costs about $0.49 an hour here (a typical cloud asks about $0.70). For 70B models, an RTX 6000 Ada or A6000 (48 GB) is the cheapest single card that fits. The full table with every size is in How much GPU memory do I need?
Not sure? Start small. A 7–8B model on a cheap card answers most everyday questions well, and trying it costs cents.
Way 1: chat in your browser (Ollama)
The Ollama environment starts a model server and a chat page you open from the rental. Pick a model and type. Behind it is Ollama, which downloads the model on the machine the first time you ask for it — the download time is not billed.
Useful model tags to start with:
llama3.1:8b— a solid all-rounder.qwen3:8bandqwen3:32b— strong at reasoning, code and many languages.deepseek-r1:14b— thinks step by step before answering.gpt-oss:20b— OpenAI's open-weight model; needs a 16 GB card.
Some hosts keep popular models downloaded in advance ("warm"), so the first answer comes in seconds instead of minutes. The marketplace shows which machines have which models ready.
Way 2: call it from your code like the OpenAI API
If you want to call a model from your own code, use the Compute API. It speaks the same language as OpenAI's API, so the OpenAI SDKs and most tools work by changing two settings: the base URL and the key (an API key you create on the developers page, paid from prepaid credits).
from openai import OpenAI
client = OpenAI(base_url="https://rentgpupower.com/api/v1/compute",
api_key="gpuapi_...", max_retries=20)
reply = client.chat.completions.create(
model="qwen3:8b",
messages=[{"role": "user", "content": "Explain transformers in two sentences."}],
)
print(reply.choices[0].message.content)If none of your machines serves that model yet, one is started for you: the first call answers "loading" while the model downloads, and the same call is served as soon as it is up (that is what max_retries is for). You pay for the GPU by the second, not per token. More in An OpenAI-compatible endpoint for open models.
On a machine you rented yourself, the vLLM environment does the same: it serves a Hugging Face model behind an OpenAI-compatible address, shown on your rental page.
Way 3: your own setup
Prefer llama.cpp, text-generation-webui, SGLang or your own code? Rent a GPU with the PyTorch + JupyterLab environment, or bring your own Docker image and a command to run when it starts. You get a terminal in the browser, and SSH if you prefer your own.
What it costs
You pay the host's hourly price, locked when you book. The money is held in escrow and released to the host only for the time the machine actually works; stop early and the rest comes back to you. Card, Apple Pay, Google Pay, PayPal and crypto are accepted.
Two examples at today's typical prices:
- An afternoon (4 hours) on an RTX 4090: about 4 × $0.49.
- An hour on a 48 GB RTX 6000 Ada for a 70B model: about $0.56.
FAQ
Can I run a 70B model on one GPU?
Yes, in 4-bit, on a card with 48 GB of memory or more (RTX 6000 Ada, A6000, A100, H100). On a 24 GB card it only runs if most of it is moved to the processor, which is very slow.
Is my conversation private?
The model runs on the host's machine inside a sandbox (a Docker container). The host cannot open your files through the app, and nothing you type is sent to an AI company. As with any rented computer, do not put secrets on it that you could not afford to lose.
How long does it take to start?
Usually a minute or two. A model that is not already on the machine is downloaded first; that time is not billed.
What happens when my time runs out?
The machine is released and your container is removed. Copy anything you want to keep to your computer or to My Drive before then — or turn on auto-save to My Drive on the rental page.