A high-throughput and memory-efficient inference and serving engine for LLMs
About this project
vllm is an open-source inference and training project on GitHub: A high-throughput and memory-efficient inference and serving engine for LLMs. It is infrastructure rather than a finished product: the layer that serves, quantises, trains or orchestrates models. It is maintained by vllm-project and published as a public repository you can read, fork and run yourself.
On GitHub the repository is tagged amd, blackwell, cuda, deepseek, deepseek-v3, gpt, gpt-oss, inference, kimi, llama, llm and llm-serving, which is a fair summary of what it covers. That generally means an API surface, batching and memory management, and often an OpenAI-compatible endpoint so existing clients keep working unchanged.
It is written mainly in Python and released under APACHE-2.0, a permissive license — commercial use is fine as long as copyright notices are kept. It is actively maintained: the most recent commit landed on November 17, 2026.
To try it, clone the repository and follow its README; that file is the canonical install guide. Because it runs on your own infrastructure, budget time for environment setup: dependencies, model weights and configuration usually take longer than the first successful run.
Key facts as recorded on GitHub: 97,101 stars; 24,929 forks; primary language Python; license APACHE-2.0; last commit on November 17, 2026; topics amd, blackwell, cuda, deepseek, deepseek-v3, gpt.