Microsoft releases Phi-3 Mini, a tiny AI model that performs on par with big models.

Microsoft releases Phi-3 Mini, a new lightweight AI model And it will be the first of three small AI models planned to be released in the future.

Phi-3 Mini has an order size of 3.8 billion parameters and trains on small datasets. Compared to large language models (LLMs) like GPT-4

The company releases Phi-3 Mini on Azure, Hugging Face, and Ollama. In addition to this one, Microsoft will release Phi-3 Small, size 7,000 million parameters, and Phi-3 Medium, size 14,000 million parameters, next.

Microsoft released the last version of Phi in December. It has capabilities comparable to larger models like the Llama 2.

The company points out that the Phi-3 performs better than previous models and will be able to respond similarly to models 10 times its size.

Eric Boyd, corporate vice president of Microsoft Azure AI Platform, said the Phi-3 Mini has the same capabilities as GPT-3.5 but in a smaller form factor.

Boyd also said Phi-3 training is done by creating a ‘study schedule’ inspired by how children learn from bedtime stories. A book that uses simple words and sentences that talk about big topics.

He points out that there aren’t many children’s books on the market. So the team created a list of more than 3,000 words and asked LLM to use it to create a children’s book to teach Phi.

Boyd says companies often find smaller models more suitable. This is because the data sets that each company has are usually small. Moreover, these models use less processing resources. And it’s much cheaper too.