Overview

• Microsoft has launched three new Phi-4 reasoning models: Phi-4-reasoning (14B parameters), Phi-4-reasoning-plus, and Phi-4-mini-reasoning (3.8B parameters).
• Phi-4-reasoning outperforms much larger models, including DeepSeek’s 671B MoE model, on math and science benchmarks like AIME 2025.
• Phi-4-mini-reasoning, designed for edge devices, surpasses models twice its size in reasoning tasks, enabling educational and offline applications.
• All models are open-weight and available on Azure AI Foundry and Hugging Face, promoting accessibility for developers and enterprises.
• Integrated into Copilot+ PCs via the Phi Silica variant, these models offer ultra-low latency AI services, even in offline environments.