Dec 14, 2024 00:00:00

Microsoft releases 'Phi-4', an AI model that is lightweight but has overwhelmingly superior mathematical performance to GPT-4o

Microsoft released the language model ' Phi-4 ' on December 13, 2024. Phi-4 has 14 billion parameters, which is small compared to major large-scale language models, but it has top-class mathematical performance.

Introducing Phi-4: Microsoft's Newest Small Language Model Specializing in Complex Reasoning | Microsoft Community Hub

https://techcommunity.microsoft.com/blog/aiplatformblog/introducing-phi-4-microsoft%E2%80%99s-newest-small-language-model-specializing-in-comple/4357090

Microsoft is developing a small-scale, high-performance language model 'Phi' series, releasing ' Phi-2 ' in December 2023 and ' Phi-3 ' in April 2024. In addition to the advanced language processing performance of the previous Phi series, the newly released Phi-4 also supports tasks that require complex reasoning such as mathematics.

In the graph below, the horizontal axis shows the number of parameters, and the vertical axis shows the score of the benchmark test ' MMLU '. Phi-4 recorded higher scores than language models with the same parameter size, and showed performance comparable to language models with more parameters such as 'Llama-3.3-70B-Instruct' and 'Qwen2.5-72B-Instruct'.

The following graph compares the average scores of various AI models when solving problems from

the American Mathematics Competition (AMC) . Phi-4 scored higher than heavyweight models such as GPT-4o and Gemini Pro 1.5.

Here is an example of how to solve a math problem with Phi-4. The correct answer is given with an explanation of how to solve it.

Phi-4 is already available on Azure AI Foundry and will also be available on Hugging Face within the next week.