Vibeleaderboard
Liquid AI

ai lab

Liquid AI

Liquid AI matters because it is testing whether model architecture, hardware-aware training, and deployment software can make useful foundation models live on personal and embedded devices. That is a different optimization target from maximizing a cloud model's benchmark score, and it creates inspectable tradeoffs among capability, latency, memory, privacy, and cost.2,6,10,4,14

Official site ↗Models in the Atlas →

Profile

Overview

An MIT research spinout

Liquid AI is a Cambridge, Massachusetts model company founded in 2023 as a spinout from the MIT Computer Science and Artificial Intelligence Laboratory. Its founders are Ramin Hasani, Mathias Lechner, Alexander Amini, and Daniela Rus. Before the company existed, those four researchers and Radu Grosu introduced liquid time constant networks for adaptive continuous-time sequence processing. Liquid AI was created to turn that broader efficiency research tradition into general-purpose systems rather than remain a robotics research project.1,2,5

Financing an architectural thesis

The company emerged from stealth in December 2023 with seed financing and a stated plan to build alternatives to conventional transformer systems. It raised a $250 million Series A led by AMD in December 2024. The financing and hardware partnership gave Liquid resources to train models and optimize them across CPUs, GPUs, and accelerators, while retaining a narrower product thesis than frontier labs built around very large cloud models.2,3

What an LFM actually is

Liquid Foundation Models are related to the founders' efficiency research but should not be described as ordinary liquid neural networks scaled into language models. The LFM2 technical report documents a hardware-guided hybrid architecture that combines gated convolutions with a smaller number of attention blocks. The family ranges from hundreds of millions to several billion parameters and includes text, vision-language, and audio variants designed for on-device use. The weights are downloadable under the LFM Open License, whose free commercial-use grant is limited to companies under $10 million in annual revenue.6,7,8

The edge deployment program

In 2026, Liquid concentrated on LFM2.5, inference formats, and the Liquid Edge AI Platform. It released compact agentic checkpoints, quantization-aware distilled variants, speculative decoding models, and a device benchmark suite. One independent Apple M4 test found real speed advantages alongside failures in a small prompt sample. Artificial Analysis separately measures model intelligence and says it independently validated Liquid's inference measurement process for their joint mobile benchmark. Liquid has demonstrated a coherent edge model program, but it has not yet shown that compact local models can replace larger cloud systems on complex reasoning and coding work.10,12,13,4,14

Notable contributions

  1. 01Liquid time constant networksBefore Liquid AI was founded, Ramin Hasani, Mathias Lechner, Alexander Amini, Daniela Rus, and Radu Grosu introduced a class of continuous-time recurrent neural networks whose time constants vary through learned nonlinear dynamics. That work established the academic line from which the later company took its name.5,2
  2. 02Hardware-guided LFM2 architecture searchLFM2 used hardware-in-the-loop architecture search to choose a compact hybrid of gated convolutions and grouped-query attention blocks against edge latency and memory constraints.6
  3. 03A multimodal on-device model familyLiquid built text, vision-language, audio, and mixture of experts variants around the LFM2 family, then paired them with local deployment tooling instead of treating edge conversion as an afterthought.7,9,10
  4. 04Model and inference co-design for local agentsThe LFM2.5 program added quantization-aware distillation, speculative decoding checkpoints, and device-level measurements to reduce latency and memory use for local tool-calling workloads.10,11,12,13