High-throughput serving engine for running open LLMs with an OpenAI-compatible API

v0.25.12026-07-14PythonApache-2.0
86.9k Completely free Open source

Doc version 1.0 · Updated

Related tools

Overview

vLLM is a high-throughput, memory-efficient inference and serving engine for LLMs, best known for PagedAttention and continuous batching.
It serves open models behind an OpenAI-compatible API, so anything that speaks OpenAI — including LiteLLM — can call your self-hosted endpoint.

flowchart LR
  model[open model] --> vllm[(vLLM server)]
  req[OpenAI-style request] --> vllm
  vllm --> resp[completion]

The Code samples tab shows serving a model and calling it over HTTP.

When to use it

Choose vLLM when you run open models on your own GPUs and need maximum throughput and concurrency — production self-hosted inference rather than a desktop or single-stream local runtime.