Overview

• Microsoft introduced three new Phi-4 reasoning models, including Phi-4-reasoning, Phi-4-reasoning-plus, and Phi-4-mini-reasoning, designed for complex reasoning tasks.
• Phi-4-reasoning and Phi-4-reasoning-plus outperform much larger models, such as DeepSeek-R1 (671B parameters), on benchmarks like AIME 2025 and Ph.D.-level science tests.
• The compact Phi-4-mini-reasoning model, with 3.8 billion parameters, matches or surpasses larger 7–8 billion-parameter models on math reasoning benchmarks.
• All models are open-weight and available on Azure AI Foundry and Hugging Face, supporting both cloud and offline edge deployment, including integration into Copilot+ PCs.
• Advanced training techniques, including reinforcement learning and synthetic data, enable these models to balance compact size with high performance, aligning with Microsoft's responsible AI principles.