Self-host vLLM

updated Aug 2026prices checked · Jul 2026

We earn commissions when you shop through the links below. Full disclosure →

High-throughput LLM inference and serving engine with PagedAttention and continuous batching — an OpenAI-compatible API server built for GPU-backed, multi-user production workloads.

Key facts

CategoryAI Chat Interfaces

LicenseApache-2.0

StackPython, CUDA

Min RAM16384 MB

Official imageyes

Difficulty

Sourcegithub.com/vllm-project/vllm

Websitevllm.ai

Our recommendation

Reach for vLLM when a GPU server needs to serve many people at once — continuous batching and PagedAttention keep one card ahead of a whole team's chat traffic, behind the same OpenAI-compatible API your clients already speak. If it's just you, or your VPS has no GPU, Ollama does the job with a tenth of the setup.

What you need

Any VPS with at least 16384 MB of RAM
A domain you control — most self-hosted setups need HTTPS in front of them
About an afternoon — budget time for troubleshooting

Where to host itaffiliate disclosure

Hetzner Cloudrun it on

From $23.59/mo · 2 vCPU / 4 GB / 80 GB · EU + US

Get Hetzner Cloud

Kamaterafree trial

From $4/mo · 1 vCPU / 1 GB / 20 GB · US + EU + Asia

Start free on Kamatera →

DigitalOceanalso works on

From $6/mo · 1 vCPU / 1 GB / 25 GB · US + EU + Asia

Deploy on DigitalOcean →

Paid link — we earn a commission if you shop through it.

Install

Run these commands on your server:

# vLLM — official image needs an NVIDIA GPU (CPU builds exist but are a separate image)
docker run -d --runtime nvidia --gpus all -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 --ipc=host vllm/vllm-openai:latest --model Qwen/Qwen3-0.6B

vps picksBest VPS for self-hostingbenchmarked quarterly · re-priced monthly

What you take on

vLLM is production-grade serving infrastructure, and it asks you to operate like it:

non-negotiableIt effectively requires a GPU, NVIDIA first. The official vllm/vllm-openai image assumes CUDA (AMD gets a separate vllm/vllm-openai-rocm image) — you'll install the drivers and the NVIDIA container toolkit before anything serves, and the CPU builds exist mostly for development, not the throughput vLLM is chosen for.

non-negotiableVRAM math comes first. vLLM loads full Hugging Face weights, so the model claims most of your card before the clever cache management helps — size the GPU to the model you actually want to serve, or startup will fail with an out-of-memory error, not a warning.

non-negotiableIt moves fast. Releases land frequently and flags shift between them — pin an exact image tag rather than latest, and re-read the server flags on every upgrade instead of assuming yours still mean the same thing.