Microsoft announced the launch of its latest AI model, Phi-3-mini, the first in a series of three smaller AI models that the company plans to release. This model, which measures 3.8 billion parameters, is now available on Azure, Hugging Face, and Ollama.

Phi-3-mini is a lightweight AI model that is trained on a smaller dataset compared to larger language models like GPT-4. Despite its smaller size, Microsoft claims that it performs better than its predecessor, Phi-2, and can provide responses that are similar to those of a model that is 10 times larger.

Eric Boyd, corporate vice president of Microsoft Azure AI Platform, described Phi-3-mini as being just as capable as larger language models like GPT-3.5, but in a smaller form factor. He added that smaller AI models like Phi-3 are often cheaper to run and perform better on personal devices like phones and laptops.

Microsoft has been focusing on building a team dedicated to developing lighter-weight AI models, and Phi-3 is the latest addition to the company’s growing list of AI models. In addition to Phi, Microsoft has also developed Orca-Math, a model that is focused on solving math problems.

Microsoft’s competitors have also been developing their own small AI models, with Google’s Gemma 2B and 7B being used for simple chatbots and language-related work, Anthropic’s Claude 3 Haiku for summarizing dense research papers, and Meta’s Llama 3 8B for chatbots and coding assistance.

Boyd explained that developers trained Phi-3 using a “curriculum” that was inspired by how children learn from bedtime stories, books with simpler words, and sentence structures that talk about larger topics. He added that Phi-3 simply built on what previous iterations had learned, with Phi-1 focusing on coding and Phi-2 beginning to learn to reason. Phi-3, on the other hand, is better at coding and reasoning.

While the Phi-3 family of models knows some general knowledge, it cannot beat a GPT-4 or another large language model in terms of breadth. However, Boyd noted that smaller models like Phi-3 are often better suited for custom applications since many companies have smaller internal datasets. Additionally, smaller models use less computing power, making them more affordable.

This article was originally published on April 23, 2024, at 3:30 am ET.