TraviaTechPie Review

The Story

Here’s the bet almost everyone in AI has been making for the last three years: bigger is better. More parameters, more data, more GPUs, more electricity. The frontier is a cloud data center the size of a warehouse, and your phone is just a window into it. Liquid AI — the MIT spinout behind the “Liquid Foundation Model” line — just placed a very different bet.

관련해서 Meta putting a 30B agent model on one GPU도 함께 참고하시면 좋습니다.

Another take on smaller, more efficient models: UNIST’s GMoE parameter-sharing.

For the infrastructure side of the agent economy, see Naïve’s agent infrastructure.

The company unveiled a family it calls “Liquid Nanos”: extremely small foundation models built to run directly on everyday devices. We’re talking 350 million to 2.6 billion parameters. For context, the models people actually use in the cloud are often measured in the hundreds of billions or trillions. Nanos are, by that yardstick, tiny. The whole point is that they’re small enough to live on a phone, a laptop, or an embedded chip — no data center required, no round trip to a server.

The initial release is six task-specific models, and the names tell you the philosophy right away. There’s LFM2-350M-ENJP-MT for English–Japanese translation. LFM2-350M-Extract and LFM2-1.2B-Extract for pulling structured data out of messy documents. LFM2-350M-Math for math problems. LFM2-1.2B-RAG for answering questions from long documents. And LFM2-1.2B-Tool for function calling — the plumbing that lets an AI actually trigger actions instead of just chatting. Notice what’s missing: there’s no “do everything” chatbot in the lineup. Each model does one job.

That’s the thesis, and it’s worth sitting with. The dominant idea in agentic AI has been to take one enormous generalist model and ask it to do everything — translate, extract, reason, call tools, hold a conversation. Liquid’s argument is that for most of the actual work an agent does, you don’t need the generalist. You need a specialist that’s been sharpened for exactly that task. And a sharpened specialist can be hundreds of times smaller.

The performance claims are where it gets interesting, and it’s worth being precise about which model is being compared to what — because the numbers are easy to misread. According to Liquid’s own evaluations, the 350M extraction model, LFM2-350M-Extract, beats Google’s Gemma 3 4B on structured extraction — a generalist roughly 11 times its size. The larger LFM2-1.2B-Extract goes further, reportedly exceeding Gemma 3 27B (around 22.5 times bigger) and rivaling GPT-4o (about 160 times bigger). To be clear, that’s the 1.2B model beating the 27B — not the 350M one. These are the company’s internal and partner evaluations, so take them with the usual grain of salt. But the direction is unambiguous: on a narrow task, a well-trained small model can punch far above its weight.

The models ship through Liquid’s own Edge AI Platform, called LEAP, and are also on Hugging Face and in an app called Liquid Apollo. They’re available under an open license for academics, developers, and small businesses (commercial terms apply for larger-scale use). So this isn’t a research teaser locked behind a waitlist — you can pull these down and run them today.

The line the company keeps coming back to is from its co-founder and CTO, Mathias Lechner: Nanos deliver “the task-specific performance of large frontier models at zero marginal inference cost.” Unpack that phrase, because it’s the whole business case. When a model runs on the user’s own device, every query after the first one is essentially free to the company serving it. No GPU rented by the hour. No electricity bill that scales with your popularity. The cost curve of AI, which right now bends sharply upward every time you add a user, flattens out.

The Takeaway

I’ve been circling this exact theme for weeks now, and Liquid just gave it a sharper edge. Back when Thinking Machines released Inkling-Small, the story wasn’t the model’s capability — it was that being a quarter the size was the whole point. Nanos are the same argument taken further: not just smaller, but small enough to leave the data center entirely.

And that’s the shift worth naming. For a while the interesting question in AI was “how big can we go?” The more interesting question now is “how small can we get away with?” Those are opposite optimization targets, and the second one has been quietly gaining ground. There’s a reason. If you’ve followed the memory-wall pieces here, you know the real constraint on giant models isn’t raw compute anymore — it’s feeding the beast: memory bandwidth, energy, the cost of every single inference. Small on-device models don’t fix that wall. They route around it.

Here’s what I find genuinely compelling about the “specialist” framing, though. It’s a quiet rebuke of a whole design pattern. The industry decided that intelligence means generality — one model that can do anything. Liquid is saying that’s often the wrong tool. Most agent workflows are boring and repetitive: extract these fields, translate this string, call that function. You don’t want a philosopher for that. You want a fast, cheap, reliable specialist that runs locally and never phones home. That’s a very different picture of what “AI infrastructure” looks like — less a cathedral in the cloud, more a swarm of small tools scattered across billions of devices.

It also changes the parts of AI that never make the headlines: privacy, latency, and offline reliability. A model on your device doesn’t send your documents to someone’s server to get parsed. It answers instantly because there’s no network hop. And it keeps working on a plane, in a tunnel, in a hospital basement. For a lot of real applications — the medical, legal, and industrial ones where data can’t leave the building — that’s not a nice-to-have. It’s the only version that’s allowed to exist.

So do I think this ends the era of giant cloud models? No. Frontier reasoning, the hardest open-ended problems, the stuff that genuinely needs a trillion parameters — that lives in the data center for a long time yet. What’s happening is a split. The big brains stay in the cloud for the hard thinking, and a fleet of tiny specialists handles the millions of small, well-defined jobs at the edge. Liquid is betting the second category is far bigger than the first, and honestly, if you count how many of an agent’s actual steps are just “extract, translate, call, retrieve,” they might be right.

The one thing I’d watch is the gap between benchmark and reality. “Beats a model 500 times its size” is true on the specific task it was tuned for — and can fall apart the moment the input drifts outside that task. A specialist is only as good as the boundary of its specialty. That’s the tradeoff the whole approach rests on, and it’s the thing developers will find out fast once these are running in the wild.

This article is for informational purposes only and is not investment or product-purchase advice.

TL;DR: Liquid AI released “Nanos” — six task-specific small language models (350M–2.6B parameters) that run on phones, laptops, and embedded devices instead of the cloud. Per the company’s own evaluations, LFM2-350M-Extract beats Gemma 3 4B (~11x larger) and the bigger LFM2-1.2B-Extract exceeds Gemma 3 27B (~22.5x) while rivaling GPT-4o on structured extraction. The bet: most agent work doesn’t need a giant generalist, just a fast, cheap specialist running locally — at “zero marginal inference cost.”

Photo: Igor Omilaev / Unsplash