LMSYS Org

Research infrastructure and inference

LMSYS Org is the public name of Large Model Systems, a nonprofit that grew out of a university collaboration. Its projects include FastChat for model serving and evaluation, Vicuna conversational-model research, and SGLang for executing language-model programs. The organization now lists Chatbot Arena as a graduated project alongside SGLang, FastChat and Vicuna. [1]

Its January 2024 SGLang account paired a Python programming interface with an inference runtime. RadixAttention reuses cached computation for shared prompt prefixes, while the frontend expresses multi-call control flow. That contribution concerns how models execute; it is distinct from releasing a new set of model weights. [2]

The later SGLang serving account described an Apache-licensed runtime for language and vision-language models. Its benchmark correction is a useful boundary on the headline results: performance depends on workload, hardware and implementation, and the authors revised their own comparison after finding a data-generation bug. [3]