SGLang is an open-source serving framework for high-performance inference and training workflows with large language and multimodal models. It is designed to make complex model programs execute efficiently while supporting modern workloads such as reasoning models, agentic systems, structured generation, multimodal inference, and large-scale model serving.

The project combines a model-serving runtime with optimization techniques for throughput and latency, including advanced scheduling, caching, speculative decoding, parallelism, quantization support, and hardware-specific acceleration. It tracks new open models closely and frequently provides early or day-zero support across NVIDIA GPUs and other accelerator platforms, while related components extend the stack to image, video, audio, and reinforcement-learning workloads.

SGLang is primarily aimed at AI infrastructure engineers, researchers, and organizations operating demanding model-serving systems. It is useful when basic inference servers are not enough and teams need high throughput, efficient resource use, flexible model support, and a runtime that can serve as infrastructure for production agents, APIs, evaluation pipelines, or large-scale research.