LFM

What it is

Liquid AI LFM is the family of models built by Liquid AI, a company spun out of MIT, and it represents one of the more deliberate attempts to build capable language models without the transformer at their core. The name stands for liquid foundation models, and the project is best understood as a bet that the dominant architecture of recent years is not the only way, or always the best way, to build a useful model. Liquid AI LFM is aimed squarely at running efficiently, and in particular at running on the edge: on phones, laptops and embedded hardware rather than in a distant data centre.

The bet against the transformer

To appreciate the bet, it helps to know what almost everyone else is doing. The overwhelming majority of today’s language models are transformers, and much of the recent progress in AI has come from scaling that single design ever larger. Liquid AI has taken a different starting point, drawing on the company’s research heritage in systems inspired by biological neural circuits, and built LFM around an alternative to attention. The goal is efficiency rather than novelty for its own sake: a model that does more with less memory and less compute, so that useful capability can fit where a large transformer cannot.

Why efficiency and the edge

Efficiency is the thread that ties the whole project together. Running a model on the edge imposes hard limits that a data centre does not. A phone has a fixed amount of memory, a battery to protect and no guarantee of a fast connection, so a model that runs there has to be economical by design rather than trimmed down as an afterthought. By treating edge deployment as the target from the start, Liquid AI LFM is shaped by those constraints, and an edge ai model that runs locally brings real advantages: responses without a round trip to a server, function without an internet connection, and data that need never leave the device.

That last point speaks to a broader shift in how the field thinks about where computation should happen. Sending every request to the cloud is convenient but carries costs in latency, in running expense and in privacy, since the user’s data has to travel to someone else’s servers to be processed. Models that run on the device sidestep those costs, and Liquid AI has positioned LFM to serve exactly that demand. Its MIT-spinout origins give the effort a research pedigree, and its focus on efficiency gives it a clear reason to exist alongside the larger, cloud-bound systems.

The case for non-transformer designs

The wider significance of Liquid AI LFM is that it keeps alive the question of whether the transformer deserves its near-monopoly. For most of the field’s recent history, the safe assumption has been that the path forward runs through bigger transformers trained on more data. Efforts like LFM test that assumption by shipping an alternative rather than merely proposing one, and by aiming it at a segment, the edge, where the transformer’s appetite for memory and compute is most obviously a handicap.

What to watch is whether efficiency-first, non-transformer models can match the quality of the systems they hope to displace on the tasks that count. Users care about what a model can do before they care about how it is built, so an alternative architecture has to earn its place on results, not on principle. If on-device capability becomes a mainstream expectation, and there are good reasons to think it will, designs built for that world from the outset stand to benefit. For readers following the range of approaches across the field, our large language models hub covers the families in play, and the broader AI section tracks the labs and the strategies behind them.

For the current LFM line-up and the technical details Liquid AI chooses to publish, the company’s own site is the authoritative source and the place to check before relying on any particular figure.