Small language models: 10 Breakthrough Technologies 2025

Large language models unleashed the power of AI. Now it’s time for more efficient AIs to take over.

Will Douglas Heavenarchive page

January 3, 2025

Selman Design

WHO

Allen Institute for Artificial Intelligence, Anthropic, Google, Meta, Microsoft, OpenAI

WHEN

Now

Make no mistake: Size matters in the AI world. When OpenAI launched GPT-3 back in 2020, it was the largest language model ever built. The firm showed that supersizing this type of model was enough to send performance through the roof. That kicked off a technology boom that has been sustained by bigger models ever since. As Noam Brown, a research scientist at OpenAI, told an audience at TEDAI San Francisco in October, “The incredible progress in AI over the past five years can be summarized in one word: scale.”

But as the marginal gains for new high-end models trail off, researchers are figuring out how to do more with less. For certain tasks, smaller models that are trained on more focused data sets can now perform just as well as larger ones—if not better. That’s a boon for businesses eager to deploy AI in a handful of specific ways. You don’t need the entire internet in your model if you’re making the same kind of request again and again.

Most big tech firms now boast fun-size versions of their flagship models for this purpose: OpenAI offers both GPT-4o and GPT-4o mini; Google DeepMind has Gemini Ultra and Gemini Nano; and Anthropic’s Claude 3 comes in three flavors: outsize Opus, midsize Sonnet, and tiny Haiku. Microsoft is pioneering a range of small language models called Phi.

A growing number of smaller companies offer small models as well. The AI startup Writer claims that its latest language model matches the performance of the largest top-tier models on many key metrics despite in some cases having just a 20th as many parameters (the values that get calculated during training and determine how a model behaves).

Explore the full 2025 list of 10 Breakthrough Technologies.

Smaller models are more efficient, making them quicker to train and run. That’s good news for anyone wanting a more affordable on-ramp. And it could be good for the climate, too: Because smaller models work with a fraction of the computer oomph required by their giant cousins, they burn less energy.

These small models also travel well: They can run right in our pockets, without needing to send requests to the cloud. Small is the next big thing.