SGLang
Open-source software for serving large language models efficiently, commonly shipped alongside vLLM as a recommended way to run open-weight releases.
SGLang is a serving framework. Downloading a model’s weights gives you a very large pile of numbers; something has to load them onto GPUs, handle many users at once, batch their requests sensibly and stream answers back. That is the job SGLang does, and it is where a lot of the practical speed of a self-hosted model comes from.
It sits in the same category as vLLM, and open-weight model releases increasingly ship recipes for both on day one, which is a quiet sign of maturity: the model is not just published, it is publishable in a form people can actually run. For anyone self-hosting, the choice between them usually comes down to which handles your particular model architecture and hardware best.