from 24 GBChat
Private ChatGPT · Qwen3 32B
A chat in your browser for writing, code and documents. No subscriptions and no message limits.
from €0.30/h · RTX 3090, 24 GB
Run AI models · vLLM
vLLM on a rented GPU: your own LLM API
An 8B model in FP16 needs about 19 GB, so a 24 GB card. 32B in AWQ takes about 23 GB, and 70B in AWQ about 50 GB on one or several cards. vLLM serves many requests at once and exposes an OpenAI-compatible API, which is why people use it for chatbots, RAG and services with real users.
- App
- vLLM
- GPU
- RTX 3060, 12 GB
- Price
- from €0.12/h
To launch, your balance must cover one hour of the machine. You are billed by the minute for actual use.
from 6 GBof video memory
2–5 minutesuntil the server is up
By the minutebilled from your balance
SSH and Jupyterserver access
How much memory you need and which GPU to pick
vLLM 8B in FP16 takes about 6 GB of video memory in 4-bit; the cheapest suitable card is RTX 3060. Prices update from available catalogue machines when you open the page.
| Version | FP16 | 4-bit | Suitable GPUGPU | Price per hourPer hour |
|---|---|---|---|---|
| vLLM 8B in FP16 | 18 GB | 6 GB | RTX 3060 | €0.12 |
| vLLM 32B in AWQ (4-bit) | 71 GB | 22 GB | 2× RTX 3060 | €0.24 |
| vLLM 70B in AWQ (4-bit) | 154 GB | 47 GB | 4× RTX 4060 Ti | €0.64 |
Estimate: FP16 ≈ 2 GB per billion parameters, 4-bit ≈ 0.6 GB, plus 10% for context. Pick: the cheapest configuration with enough memory; for large models that means several identical cards. All GPUs and prices
How to run vLLM in 3 steps
- Sign up and top up. Create an account and top up from €10. To launch, your balance must cover at least one hour of the chosen machine; after that you pay by the minute.
- Choose a GPU. Memory for each version is in the table above, and the lightest one runs from 6 GB. The button on this page selects the cheapest option with the vLLM template.
- Run the model. The server is ready in 2–5 minutes. Open Jupyter in your browser or connect over SSH and run the commands below.
Finished? Stop the server in the dashboard and billing stops. Save and stop keeps /workspace for €0.03 per GB per month, and it comes back on your next server automatically.
Start the server
pip install vllm
vllm serve Qwen/Qwen2.5-7B-Instruct --host 0.0.0.0 --port 8000
# API: http://<address>:8000/v1/chat/completions
What people run vLLM for
01
A chatbot under real load
Dozens of users at once: vLLM batches their requests and keeps response times steady.
02
RAG and agents
Your own /v1/chat/completions endpoint for LangChain, LlamaIndex and any OpenAI SDK client.
03
Batch processing
Labelling, translating and summarising thousands of documents in one rental, instead of paying per token.
Tips before you launch
- The --max-model-len option limits the context and saves memory.
- For 70B on two cards, use --tensor-parallel-size 2.
Just need a chat? A ready-made app with no setup
If you need the model for conversations and don’t want to write code, launch Private ChatGPT: the server comes with Ollama, a Qwen3 model and the Open WebUI chat window in your browser.
Frequently asked questions
Is vLLM compatible with the OpenAI SDK?
Yes. Set base_url to your server and your client code stays the same.
How long does a launch take?
The server is usually ready in 2–5 minutes, then SSH and Jupyter are available. Downloading the model depends on its size and is billed as normal server time.
What happens to my models and results after I stop?
When you stop, click Save and stop: the /workspace folder with your models, LoRAs and results goes to storage for €0.03 per GB per month and comes back on your next server automatically. If you stop without saving, the disk is deleted.
How is it billed?
By the minute from your balance in euros. Top up from €10; to launch, your balance must cover at least one hour of the chosen machine. Stop the server and billing stops.
A server is usually ready in 2–5 minutes
Sign up in a minute, top up from €10, pay by the minute. Stop the server and billing stops.