What are small language models, and why are they starting to rival giant AI systems?

There's long been a dominant assumption in the AI world: the more parameters a model has, the more capable it is. Liquid AI's new LFM2.5 model stands out as an example that challenges that assumption, and it's quickly drawing interest among developers working in hardware-constrained environments.

Parameters are the adjustable numerical values an AI model learns during training, used to encode language patterns, relationships and knowledge. As a rule of thumb, more parameters mean more "capacity" — but that relationship isn't always straightforward; research in recent years has shown that training data quality can matter far more than raw size.

With just 2.6 billion parameters, LFM2.5 delivers performance comparable to models roughly four times its size, around 10 billion parameters. It's the latest example of a fast-growing category known as small language models, or SLMs — a category that has gained significant momentum over the past year in both research and commercial deployment.

Behind this performance gain lies less about raw parameter count and more about training data quality, architectural optimizations, and a technique called distillation — where a small model is trained under the "guidance" of a larger one, compressing its knowledge into a more compact form, a process aimed at capturing the essence of a larger model's "wisdom" with far fewer parameters.

The biggest practical advantage of small models is that the hardware required to run them is far more modest. Large models typically run in the cloud on powerful server clusters, while models like LFM2.5 can run directly on-device — on phones, wearables, smart home systems and even robots — substantially cutting down on response latency.

That also makes a meaningful difference for privacy: data can be processed on-device without being sent to cloud servers. The ability to run without an internet connection is also a significant advantage compared with cloud-dependent large models — a distinction that matters especially for applications handling sensitive information like health data or personal communications.

Cost is another major factor: running small models is far cheaper than running massive cloud-based ones, allowing developers to build AI features into their products without racking up large cloud bills; some developers report that switching to small models has cut their operating costs by as much as tenfold.

Small models do have limits, of course: for tasks requiring very broad general knowledge or complex chains of reasoning, large models still hold the edge. SLMs shine on narrower, well-defined tasks instead — a customer service bot or a smart home assistant tailored to a specific domain, for instance.

The industry trend points toward proliferation of task-specific, efficient, on-device small models rather than a single giant "do-everything" model. It's a shift that can also be read as AI "migrating" from cloud servers to phones and wearables — one many analysts view as the AI industry's next major evolution.

Models like LFM2.5 suggest AI's future may be shaped not just by building ever-larger models, but by delivering the same capability with fewer resources, on more accessible devices.