On this page
Your own GPU, your own model
With dedicated instances you run an open model on a GPU that is yours alone. We deploy vLLM, run readiness probes and hand you an OpenAI- and Anthropic-compatible URL. Ideal when you want predictable performance or full isolation.
How it works
- Pick a model from the Model Garden or a HuggingFace ID
- Set the VRAM and choose a European GPU
- We deploy vLLM and run automatic readiness probes
- You get a warm endpoint plus an API key
New instances run vLLM with their own encrypted upstream key, so your endpoint stays private.
Pay only for what you run
Instances stay warm while someone is online and idle down when nobody is using them, so you do not pay for an unused GPU. Billed per hour on one prepaid credit balance.
Example in code
Once your instance is warm, you call it through the Router by its model slug, exactly as you would call an OpenAI model:
from openai import OpenAI
client = OpenAI(
base_url="https://hostyourai.com/api/v1",
api_key="hyai-...",
)
resp = client.chat.completions.create(
model="your-instance-model",
messages=[{"role": "user", "content": "Hello"}],
)
print(resp.choices[0].message.content)Want to deploy step by step? Read the guide Deploy your own LLM with vLLM.
Which GPU for which model
| GPU class | VRAM | Typical use |
|---|---|---|
| L40S / RTX 4090 | 24 to 48 GB | Models up to roughly 20B parameters |
| A100 / H100 | 80 GB | Models up to roughly 70B, quantised |
| 2× H100 | 160 GB | 70B at full precision or high throughput |
| 4× H100 | 320 GB | Large models and heavy production workloads |
On deploy we enforce a VRAM floor per model, so you cannot pick a GPU the model does not fit on. Live hourly prices show before you deploy and on the pricing page.
Questions about dedicated instances
Can I run my own or fine-tuned model?
Yes. Paste a HuggingFace ID, including a private repo with your own weights, and we deploy it through vLLM on your GPU.
Is my endpoint private?
Yes. New instances run vLLM with their own encrypted API key, so only your platform key can reach the endpoint.
Do I keep paying when nobody uses the model?
Instances idle down when nobody is online, so you do not pay for an unused GPU. You are billed per hour from your prepaid balance.
When do I pick dedicated over the Router?
For steady load, custom models, stricter isolation requirements or full control over behaviour. For variable traffic the shared Router is cheaper and faster to set up.
Can I link a knowledge base?
Yes. Link a knowledge base to your instance and every chat call automatically receives relevant context from your own documents. See the guide Build RAG on EU GPUs.
Where are the GPUs?
In European datacenters. Your data never leaves the EU and a Data Processing Agreement is available. For sector-specific requirements see healthcare and government.
