vLLM is an open-source inference and serving engine optimized for running large language and multimodal models efficiently at scale. It focuses on high-throughput, memory-efficient generation and provides an API server that can expose compatible models to applications without requiring developers to build their own serving stack.

The engine supports a wide range of transformer, mixture-of-experts, multimodal, embedding, classification, and related model types, with features aimed at production workloads such as batching and distributed execution. vLLM is useful for teams that self-host models and need better GPU utilization, lower serving overhead, and a scalable backend for chat applications, agents, APIs, or other systems that generate large volumes of model inference.