Liquid AI

Liquid AI was spun out of MIT CSAIL by four founders, Ramin Hasani, Mathias Lechner, Alexander Amini and Daniela Rus, and lists 2023 as its year of establishment.¹

The company is headquartered in Cambridge, Massachusetts.^2, 3

It states its mission as building efficient general-purpose AI at every scale.¹

On 30 September 2024 it announced its first series of Liquid Foundation Models, with 1B, 3B and 40B variants.⁵

That series comprised a dense 1.3B model, a dense 3.1B model and a 40.3B mixture-of-experts model.⁵

On 10 July 2025 the company released LFM2, a new class of Liquid Foundation Models designed for on-device use.¹⁰

On 5 January 2026 it announced the LFM2.5 family, built on the LFM2 architecture, with open-weight models published on Hugging Face.¹¹

In October 2026 the company introduced d1, which it calls its first decision model; it returns a probability for each possible answer without generating any tokens.¹²

Its model catalogue is organized into text, vision, audio, nano and decision categories.⁶

Liquid AI offers its models through an API, as licensable open-weight models and through custom deployments.⁷

The LFM weights are free to use commercially until a company passes 10 million dollars in annual revenue, and the hosted services are operated in the United States.^{3, 6, 7, 8, 9}

As of April 2026, Ramin Hasani is its co-founder and CEO.^1, 4