Skip to content

Pick your style: Denkbaar

Book a conversation

Articles

History

Phi-3: Microsoft shows that small also works

·

Microsoft released Phi-3-mini, an open language model with 3.8 billion parameters that runs on a laptop and handles up to 128,000 tokens of text.

A worker sorts forms on a laptop with a small purple chip.
Phi-3 showed that a small model on a laptop can do a lot. Illustration: Denkbaar.

What happened

On 23 April 2024 Microsoft announced Phi-3, a family of open small language models. The first model was Phi-3-mini, with 3.8 billion parameters. It was available at once on Azure, Hugging Face and Ollama. Ollama is a program to run models on your own computer. Phi-3-mini came in two versions: one for 4,000 and one for 128,000 tokens of context. Context is the amount of text the model can read at once. Larger siblings, Phi-3-small (7B) and Phi-3-medium (14B), followed later.

Microsoft said the models score better than models of the same size and the next size up, on tests for language, reasoning, code and maths.

A laptop stands beside a small and a large stack of documents.
Phi-3-mini appeared with two different context sizes. Illustration: Denkbaar.

Why it mattered then

Most attention went to ever larger models. Phi-3 showed that a small model with well-chosen training data can do a lot. It runs on a laptop, and according to Microsoft even on mobile hardware.

Hands compare a form with separate data blocks and a short block on a laptop.
Small models suit fixed tasks, such as taking data from forms. Illustration: Denkbaar.

What it means now

A small model suits fixed, simple tasks: sorting text, taking data out of a form, a short summary. It works without internet and the data does not leave your computer. For hard questions or long reasoning a large model is still better. So always test on your own examples before you choose.

Sources

  1. Introducing Phi-3: Redefining what's possible with SLMs (Microsoft Azure Blog)

Topics: open-weightsmicrosoftsmall-modelslocal