Microsoft's Phi-3 mini: Understanding the AI model

Recently, Meta launched the Llama 3 Large Language Model (LLM). Now, on April 23, Meta successfully brought forward its newest version of an AI model that is "lightweight". Yes, we are talking about the Phi-3-Mini. Microsoft calls this Phi-3 is a family of the open AI models. Microsoft says that these open AI models are cost-effective and the most capable small language models (SLMs) present so far.

Here, we explain all that we know about Phi-3-mini so far.

Understanding the Phi-3-mini

It is seen as the very first of the three small models that are expected to be released by Microsoft. Reportedly, the Phi-3-Mini has successfully outperformed the models that come in the same size. It has also performed well in several areas including those of math, coding, reasoning, and language.

It is important to understand that language models are actually trained on the data already existing to tackle common language problems like answering questions, text classification, text generation, and more.

Many get confused with the meaning of the term "Large" in the LLMs. This term actually holds two meanings. It talks about the huge size of training data along with the parameter count. Now, when talking in terms of machine learning, the knowledge and memory of the machine are the key parameters. A machine learns various thigs themselves under model training, and these parameters prove to be important. These parameters determine how skilled a model is in terms of solving a particular problem.

The newest Microsoft model makes it possible for the customers to choose from a wide array of high end language models. Microsoft's Phi-3-mini is actually a 3.8B language model that is available on platforms like HuggingFace, Ollama, and Microsoft Azure AI Studio.

Now, what makes Microsoft's Phi-3-mini model an important one among many is the context window. The context window of an AI refers to the amount of conversation that can be written and read by it at any specific time. Its is measured in tokens. The Phi-3-mino by Microsoft is present in two variants. One of them comes with 128K tokens and the other in 4K tokens.

Longer context windows suggest that models are able to take in and reason over large text content like web pages, code, documents, and more.