Microsoft Launches New Phi-4 AI Models
Microsoft has released three new AI models in its “Phi” series, focused on compact performance with advanced reasoning capabilities. These models aim to balance high performance and efficiency, ideal for use on mobile devices and low-latency systems.
Small But Smart
The new models — Phi-4-reasoning, Phi-4-reasoning-plus, and Phi-4-mini-reasoning — bring a new class of small language models. Phi-4-reasoning is a 14-billion parameter open-weight model built to rival larger AI models on complex tasks. The enhanced Phi-4-reasoning-plus version uses reinforcement learning to boost accuracy with longer inputs, requiring more compute but offering better output quality.
Meanwhile, the Phi-4-mini-reasoning model is optimized for mathematical reasoning and runs on small devices. At just 3.8 billion parameters, it targets educational applications that need advanced performance in a compact package.
Also read: Microsoft May Host Elon Musk’s Grok AI on Azure Amid OpenAI Tensions
Trained on Quality Data and Synthetic Problems
Microsoft developed these models using distillation, reinforcement learning, and curated datasets from web sources and existing AI models. The mini version was fine-tuned with synthetic data, featuring over 1 million math problems. These problems were generated by Deepseek-R1 and spanned school to Ph.D. difficulty levels.
Training used a “teacher AI” approach, creating step-by-step solutions. This lets the model learn how to solve problems, not just memorize answers. The goal is to teach structured reasoning and problem-solving strategies, especially in math, physics, and science.
Outperforms Much Larger AI Models
Despite being smaller, the new models reportedly outperform OpenAI’s o1-mini and DeepSeek1-Distill-Llama-70B in science and math reasoning tasks. Microsoft claims its Phi-4-reasoning models even surpass the 671-billion-parameter DeepSeek-R1 on the AIME 2025 test, a high-level mathematics competition qualifier.
All three models, Phi-4-reasoning, reasoning-plus, and mini-reasoning, are now available on Azure AI Foundry and HuggingFace.
