vLLM: your own LLM server
OpenAI-compatible API for Llama, Qwen, Mistral and more
- Video memory
- 12 GB
- Setup
- ~6 min
- Access
- port 8000
- Billing
- hourly, in BRL
vLLM is the most widely used inference server for putting an open model into production: OpenAI-compatible API, continuous batching and efficient use of video memory. Swap the base URL in your app and it starts talking to your own model.
vLLM is the fastest LLM server (PagedAttention). Starts with Qwen2.5-3B-Instruct (multilingual, great PT-BR) — edit compose to switch to any HF model. /v1/chat/completions plug-and-play with any OpenAI client.
What it is for
- Replace per-token API spend with a fixed hourly cost
- Serve many concurrent users with continuous batching
- Keep prompts and responses inside your own infrastructure
- Try several open models without changing client code
How to deploy
- Create your account and add balance (card or Pix, no subscription).
- In the console, pick the vLLM template and a machine — the console hides the ones that do not meet the requirement.
- In about 6 minutes the setup finishes and the access address shows up in the panel, on port 8000.
Done? Just destroy the machine and billing stops with it. No contract, no minimum commitment.
FAQ
Which model should I pick?
Models in the 7–14 billion parameter range fit on 24 GB cards and cover most use cases. Above that, look at 80 GB machines.
Is the API really compatible?
Yes, the chat and completions endpoints follow the OpenAI shape, so existing libraries work by changing the base URL.
When is it cheaper than per-token?
When usage is steady. A machine running 8 hours a day with real volume usually lands well below per-token pricing; for occasional use, a token API is simpler.