Phi-4-mini, Microsoft's new next-gen small model, has finally arrived

Microsoft is back with two new models of the Phi-4 family, the Phi-4 mini and the Phi-4-multimodal.

As for the Phi-4-mini, a mini version of a family model that’s already mini, it is a compact, efficient model with 3.8 billion parameters, designed for speed and performance in text-based tasks.

Despite being smaller, it “outperforms many larger models” in tasks like reasoning, math, coding, and following instructions. The model “delivers high accuracy and scalability” for AI applications that are fast and efficient on smartphones, PCs, and even cars.

Phi-4-mini uses a dense transformer with “grouped-query attention” and can process sequences up to 128,000 tokens. It can also perform tasks like “accessing external knowledge and functionality” via APIs, which makes it great for edge AI applications in industries like healthcare and retail, where “edge AI demonstrates outstanding performance.”

Microsoft says that Phi-4-mini is outperforming many competitor AI models across several important benchmarks. The graph shows Phi-4-mini scoring especially well on math problems in the GSM8K test, where it beats all other models with a score of about 88%.

It also does really well on coding tests (HumanEval) and reasoning tasks (BigBench Hard CoT). While some models like Llama-3.1-8B and Llama-3.2-3B do better on certain tests, Phi-4-mini maintains strong scores across all seven benchmark categories.

As for the Phi-4-multimodal, which excels in processing text, images, and speech, Microsoft says that it’s slightly superior to Google’s new Gemini 2.0 Flash in several key benchmarks. The latter model now powers Google’s namesake AI chatbot.

It will also power Copilot+ PCs. Microsoft says that its lineup of AI PCs will “build upon Phi-4-multimodal’s capabilities, delivering the power of Microsoft’s advanced SLMs without the energy drain.”