THENEW_DEFAULT
Back to building
Why Build Small Language Models?31:47
BUILDINGFILM

Why Build Small Language Models?

Small language models are up to 1,000x smaller than frontier AI like GPT-5.2 and Claude Opus. That's small enough to run on-device. Learn how Maxime Labonne is building SLMs at Liquid AI.

Sep 3, 2026 31:47

Maxime Labonne

Expert Insights

A year ago, a 1-billion-parameter model couldn't reliably follow a basic instruction. Today, Maxime Labonne's team at Liquid AI ships a 1.2B model that works as a general-purpose assistant. It’s small enough to run entirely on a phone or a pair of smart glasses, with no cloud connection required.

As Head of Post-Training at Liquid AI, an MIT spinout that's raised $250M and shipped 26 open-source models in under a year, Maxime has spent his career on the opposite bet from the rest of the industry. Instead of chasing ever-larger frontier models, Liquid builds small, efficient ones from first principles.

Hear him outline:

    • Why smart glasses and other devices can't rely exclusively on cloud AI

    • Why making a model “dumber” on purpose can be the better engineering choice for a narrow task

    • Why benchmark scores don’t reflect real-world usage

    • How the Llama 4 controversy showed how easy benchmarks are to manipulate

    • Why using the same data pipeline for evaluation and training makes your results biased

    • How Liquid decides what to open-source, and why the license stays free below $10M in yearly revenue

    • What car manufacturers, Shopify, and hackathon builders are doing with small models today

If you ask ChatGPT about a stop sign, it will tell you the entire history of the sign. Their model just said: It's a stop sign. You should stop.
— Maxime Labonne, Head of Post-Training at Liquid AI

Monterail Team Analysis

    • When a model runs on-device and not via API, cost stops scaling with usage

    • Liquid's $10M revenue threshold for free licensing is a calculated bet: it captures the long tail of builders and research users while still monetizing at scale.

    • The Llama 4 checkpoint-swap story is a cautionary tale for any team leaning on public leaderboards to validate a model choice

    • Maxime's point about reusing one data pipeline for both evals and training is a warning for ML teams: without a separate process, you risk grading your own homework

    • The OEM angle signals a distinct go-to-market from typical AI SaaS

    • The hackathon projects mentioned show that a small model made for one narrow job can do it better than a big AI model