LMSYS Org

What LMSYS is

LMSYS stands for Large Model Systems (LMSYS Corp.). It is a 501(c)(3) non-profit — a US designation for a charitable organisation that has no shareholders and is expected to serve a public purpose rather than return a profit. LMSYS describes its own role as incubating open-source projects and research^[1]. Its stated mission is to make large AI models accessible to everyone by co-developing open models, datasets, systems and evaluation tools^[1]. That sentence is a fair summary of how the organisation divides its work: it is not only about training models, but about the data, the software infrastructure and the benchmarks that sit around them.

Where it came from

The group began in 2023 as a multi-university collaboration involving UC Berkeley, Stanford, UCSD, CMU and MBZUAI^[1]. In September 2024 it was incorporated as a non-profit, with the stated purpose of incubating early-stage open-source and research projects^[1].

The move from a loose collaboration of researchers to a legal entity has a practical meaning. A research project usually lives inside a lab or a single student's tenure; an organisation can hold a project, accept donations and sponsorship, and keep it running after the original authors move on. The word "incubate" in the organisation's own description points at exactly this: projects start small, and the organisation's job is to give them room to grow or to be handed off.

What it builds

LMSYS lists four flagship projects: Chatbot Arena, SGLang, FastChat and Vicuna. Chatbot Arena is marked as graduated in that list^[1].

The organisation says its projects combine open models with millions of downloads and crowdsourced platforms with millions of users^[1]. It also says it trains large language models for broad accessibility and builds distributed systems to optimise their training and inference^[1]. Those two halves explain the name. The "large model" half is the models themselves; the "systems" half is the engineering that makes them trainable and servable at scale — the part of the work that is usually invisible but determines whether a model can actually be used by anyone outside the lab that trained it.

Who is behind it

The officers listed on its page are Ying Sheng, Lianmin Zheng, Mingyi Lu, Jerry Zhou and Banghua Zhu^[1]. Its advisors are Joseph E. Gonzalez, Ion Stoica, Eric P. Xing and Hao Zhang^[1].

Sponsors and donors listed by the organisation include NVIDIA, Google Cloud, AMD, Kaggle, MBZUAI, Together, Anyscale, Hugging Face and Voltage Park^[1]. The mix is notable: chip makers, cloud providers, a data-science competition platform, a university, and several companies that host or serve open models. For a non-profit that gives away models and runs public evaluation platforms, that kind of support is what pays for compute and hosting.

Where it sits

The organisation occupies an unusual position in the AI landscape. Most frontier models are built inside companies and released with limits on how they can be used; most academic work is published but rarely maintained as usable software. LMSYS sits between the two, treating models, datasets, systems and evaluation tools as one package and releasing them openly^[1].

The most visible consequence is that its projects are used both by researchers who want a reference implementation and by ordinary users who simply want to try a model. The organisation's own framing — open models with millions of downloads, crowdsourced platforms with millions of users^[1] — describes a two-way relationship: the public uses the tools, and the public's use is itself part of what the tools measure.

Its incorporation in September 2024 marked the point at which this way of working was given a permanent institutional form, with officers, advisors, sponsors and a stated mission rather than a changing cast of collaborating labs^[1].