Microsoft launches mini AI model

Microsoft has released a mini AI model, called Phi-3 Mini (available on Microsoft Azure AI Studio, Hugging Face, and the Ollama framework), which is the first–in a family of lightweight models it plans to release–designed to bring AI capabilities to smartphone devices.

Why is Phi-3 Mini smaller?

Phi-3 Mini has 3.8B parameters, which is significantly less than models like GPT-4 and Llama 3, which are rumored to have around 1 trillion parameters.

What does this mean?

In AI, parameters relate to how AI models tackle complex tasks: The more parameters a language model has, the bigger it is and the more computing power it needs to run. But the more parameters a model has, also means the more capable it is of handling nuanced requests, and the better it is at grasping context and understanding natural language etc...

But smaller models, Phi-3 Mini, aren’t designed for complex tasks. They’re designed to handle simpler, everyday tasks, tasks like drafting emails, looking for local restaurants, data analysis, document summarisation, and basic chatbot interactions.

So, because it’s smaller, the Phi-3 Mini doesn’t need the huge amount of expensive and carbon-heavy computational power that the larger LLMs need, so it can run on smartphones at a cheaper cost, with less damage to the environment.

How does Phi-3 Mini perform compared to other, larger language models?

Although it’s smaller and more lightweight, the Phi-3 Mini matches the capabilities of larger language models: For instance, it beat Llama 3 on natural language understanding and surpassed both GPT-3.5 and Llama 3 on arithmetic reasoning.

Although it scored lower on trivia and factual knowledge, researchers believe that these ‘’weaknesses’’ will be resolved once it's integrated with a search engine.

How is Phi-3 Mini able to match larger model capability?

Phi-3 Mini has been trained on high-quality educational web and synthetic data. Researchers prioritized data quality over quantity, so it can run on fewer parameters, whilst retaining accuracy.

As well as high-quality data, it’s also been trained on bedtime stories, with researchers taking inspiration from how children learn by reading stories with simpler language.