Microsoft has launched the next version of its lightweight AI model Phi-3 Mini, the first of three smaller models the company plans to release.
Phi-3 Mini measures 3.8 billion parameters and is trained on a relatively small dataset. Large language models like GPT-4. Currently available on Azure, Hugging Face, and Ollama. Microsoft plans to release Phi-3 Small (parameter 7B) and Phi-3 Medium (parameter 14B). Parameter refers to the number of complex instructions that the model can understand.
company Phi-2 released in DecemberMicrosoft says Phi-3 performs better than previous versions and can provide responses closer to 10 times larger models.
Eric Boyd, corporate vice president of Microsoft Azure AI Platform, said: The Verge Phi-3 Mini provides functionality equivalent to LLMs like GPT-3.5 „in a smaller form factor.“
Smaller AI models compared to larger AI models Personal use is often cheaper to run and has better performance devices such as phones and laptops. information Earlier this year, Microsoft Build a team specifically focused on lightweight AI models. The company, together with Phi, killer whale mathA model that focuses on solving mathematical problems.
Boyd said the developers trained Phi-3 with a „curriculum.“ They were inspired by how children learn from bedtime stories, books with simpler words, and sentence structures that talk about larger topics.
“There aren’t enough children’s books out there, so we took a list of over 3,000 words and asked LLM to create a ‘children’s book’ to teach Phi,” Boyd said. Masu.
He added that Phi-3 simply builds on what was learned in previous iterations. Phi-1 focused on coding, Phi-2 started learning reasoning, but Phi-3 is better at coding and reasoning. The Phi-3 model family knows some general knowledge, but in terms of breadth he cannot beat GPT-4 and his other LLMs. There's a big difference between the kinds of answers you'd get from an LLM trained on the entire Internet and the kinds of answers you'd get from an LLM trained on the entire Internet. A small model like Phi-3.
Boyd notes that many companies have small internal datasets anyway, so smaller models like Phi-3 are often better suited for custom applications. These models also use less computing power, so they are often much more affordable.