Rethinking the AI Playbook

For the past several years, the artificial intelligence industry has operated under a single, expensive assumption: bigger is always better. Tech giants have poured ratusan billions of dollars into building massive, power-hungry data centers to run monolithic models based on the Transformer architecture—the technology introduced in 2017 that powers systems like ChatGPT, Claude, and Gemini.

Founded in 2023 as an MIT spin-off by Ramin Hasani (CEO), Mathias Lechner (CTO), Alexander Amini (CSO), and MIT CSAIL Director Daniela Rus, Liquid AI is challenging that orthodoxy from first principles. Rather than chasing ever-larger cloud clusters, Liquid AI’s mission is to build efficiency-first, general-purpose AI at every scale. Their core thesis is simple yet radical: true intelligence should not be rented from a massive corporate cloud; it should be ubiquitous, private, and capable of running directly on local hardware—from smartphones and laptops to cars and miniature robots.

Philosophy and Vision: Why "Liquid"?

Liquid AI’s philosophy rests on three foundational pillars that make it one of the most compelling companies in the modern AI landscape:

  • Ubiquitous & Edge-Native Intelligence: Nearly 2.2 billion devices capable of hosting AI models—phones, PCs, wearables, and vehicles—ship globally every year. Liquid AI designs models small and light enough to live directly on these devices, eliminating cloud latency, slashing energy costs, and allowing users and enterprises to own their data and decision-making privately.
  • Biological Inspiration: While traditional neural networks are loosely inspired by the human brain, they use hundreds of billions of artificial neurons that freeze once training ends. Liquid AI’s research originated from studying the microscopic nematode worm Caenorhabditis elegans (C. elegans), which possesses only 302 neurons yet can navigate complex environments, hunt for food, and adapt to changing physical conditions in real time.
  • White-Box Explainability: Most giant AI models are "black boxes"—even their creators cannot fully trace why a specific neuron fired to produce an answer. Because Liquid AI uses drastically smaller, mathematically structured dynamical systems, its models are far easier to inspect, debug, and trust in safety-critical fields like aviation, robotics, biotech, and finance.

The Science Made Simple: How Liquid Foundation Models Work

To understand why Liquid AI is scientifically fascinating, consider how standard AI processes information versus how a Liquid system operates.

  1. From Static Weights to Continuous-Time Dynamics

In a standard Transformer model, once training is finished, the model’s parameters (weights) are frozen. Furthermore, as you feed a Transformer a longer document or video, its memory footprint—known as the KV-Cache—grows linearly or quadratically, quickly overwhelming the RAM of a normal laptop or phone.

Liquid AI originated from Liquid Neural Networks (LNNs) and Closed-form Continuous-time (CfC) networks. Instead of treating data as static snapshots, these networks model information as a continuous-time dynamical system governed by ordinary differential equations (ODEs):

In plain English: the system's internal state x(t) flows and adapts dynamically based on the incoming data I(t) and a flexible time-constant \tau. Rather than using slow numerical solvers to calculate these equations step-by-step, Hasani and his team discovered a closed-form mathematical solution, giving them the adaptability of continuous physics with the lightning-fast speed of standard deep learning.

  1. The LFM3 Tri-Hybrid Architecture (Late 2026 Breakthrough)

As Liquid AI scaled from robotics into full-scale language and multimodal intelligence, they evolved their math into Liquid Foundation Models (LFMs). With the launch of the LFM3 family in October 2026, Liquid AI introduced a groundbreaking Tri-Hybrid Architecture designed through hardware-in-the-loop automated search.

Instead of relying purely on heavy Transformer attention layers, the LFM3-1.7B backbone blends three distinct mathematical operators across its 24 layers:

  • 13 Gated Short Convolutions (kernel size 7): Rapidly capture local patterns and sequential structures with minimal memory overhead.
  • 5 Sliding-Window Attention Layers (256-token window): Maintain crisp short-term context without inflating memory.
  • 6 Global Attention Layers with Head-Wise Gating & Partial RoPE: Capture deep, long-range relationships across a massive 256,000-token context window while suppressing activation outliers so the model can be compressed to 4-bit precision (~1 GB of RAM) without losing intelligence.

Using this shared 2-billion-parameter class backbone, LFM3 unifies text (LFM3-1.7B), vision (LFM3-VL-2B), audio (LFM3-Audio-2B), full multimodal (LFM3-Omni-2B), and calibrated decision-making (d1-2B) onto a single architecture that runs natively on a Samsung Galaxy S26 Ultra, a Raspberry Pi 5, or a standard laptop.