AI Engineer
← All organizations

Foundation models and on-device AI

Liquid AI

Liquid AI builds Liquid Foundation Models (LFMs), general-purpose AI models designed for the latency, memory and privacy constraints of running on phones, laptops and other devices. Its LEAP SDK lets developers fine-tune models, prepare them for different runtimes and deploy them on chosen hardware. The company also works with enterprises on specialized applications: a multi-year Shopify licensing agreement announced in 2025 includes a production search model that completes searches in under 20 milliseconds.

An MIT CSAIL spinout, Liquid AI was founded by Ramin Hasani, its CEO, alongside Mathias Lechner, CTO, Alexander Amini, CSO, and Daniela Rus. Its research connects model design with deployment constraints. The company’s STAR framework uses evolutionary algorithms to discover neural network architectures tailored to tasks and hardware. By encoding architectures as numerical genomes, evaluating candidates and recombining successful designs, STAR can jointly optimize model quality, parameter count, cache size and latency on target hardware.

As of August 2026, the company reported 42.2 million model downloads, 56 LFMs shipped and more than 3,300 variants. Liquid AI raised a $250 million Series A in 2024 to expand compute infrastructure, develop its models and accelerate inference and fine-tuning capabilities for edge and on-premise deployment.

www.liquid.ai ↗

2 talks

Newest first

1 speaker at AIE

Affiliations reflect their AIE appearances, not necessarily current employment.

Maxime LabonneHead of Post-Training

Start here

  1. Everything I Learned Training Frontier Small Models

    Start here to learn about small-model scaling, reasoning-model doom loops, and Python-based workarounds for long-context tasks.

    Maxime LabonneAI Engineer Europe 2026

  2. Everything you need to know about Finetuning and Merging LLMs

    Start here for a comparison of fine-tuning tools and an introduction to adapter-based training and four-bit quantization.

    Maxime LabonneAI Engineer World's Fair 2024

Messages from the stage

Architecture under device constraints

Labonne examines embedding overhead and LFM2's combination of gated short convolutions and grouped-query attention, alongside inference on CPUs, smartphones, and GPUs.

Data and evaluation in post-training

The fine-tuning session connects synthetic dataset construction, diversity, and quality filtering with preference optimization and merged-model evaluation.

Affiliations reflect each recorded session, not necessarily current employment.

Company sources · checked 2026-08-27