vLLM

High-throughput serving engine for open-weight language models

AI Models & PlatformsFree planOverseasβ˜…β˜…β˜…β˜…β˜† 4.0

What is vLLM?

vLLM is an open-source library for serving large language models efficiently. Its core technique, PagedAttention, manages the key-value cache in blocks so memory is not wasted and many requests can be batched together, which raises throughput well above naive implementations. You start a server with a model name and get an OpenAI-compatible HTTP endpoint. Teams use it to self-host models behind their own applications on their own GPUs. The differentiator is throughput per GPU: it is the usual choice when serving open models at volume.

Last updated: 2026-09-20. This site only provides an index; for exact features, pricing, and licensing, see the official website.

Key features

  • PagedAttention for efficient key-value cache memory use
  • Continuous batching to raise throughput under load
  • OpenAI-compatible server with streaming responses
  • Tensor parallelism to split large models across GPUs
  • Quantisation support including FP8, AWQ and GPTQ
  • Python API for offline batch inference jobs on your own data

Pros & cons

Strengths

  • Among the fastest ways to serve open models on your own GPUs
  • Open source with rapid support for newly released models
  • Drop-in OpenAI API keeps client code unchanged

Watch out for

  • You supply and operate the GPU hardware yourself
  • Configuration options are numerous and worth tuning
  • Very large models still need multi-GPU setups

Best for & use cases

self-hosted serving, high-throughput inference, batch generation, private model endpoints

If you're comparing similar products, check the alternatives below, or browse all tools in the AI Models & Platforms category.

FAQ

What is PagedAttention?

It stores the model's key-value cache in small blocks that can be shared and reused instead of one large contiguous buffer, which cuts wasted memory and lets far more requests run at once.

Do I need multiple GPUs?

Not for small and mid-size models. Models that do not fit in one GPU's memory can be split across several with tensor parallelism, which vLLM supports through a configuration flag.

How does it compare with llama.cpp?

vLLM targets GPU servers and high concurrent throughput, while llama.cpp targets broad hardware including CPUs and small devices. Choose based on where and how heavily you need to serve.