Phi-4 Reasoning introduces a groundbreaking suite of small language models (SLMs) optimized for advanced reasoning tasks in mathematics, science, and coding. These models—Phi-4-reasoning, Phi-4-reasoning-plus, and Phi-4-mini-reasoning—leverage inference-time scaling and multi-step decomposition to deliver capabilities traditionally associated with larger frontier models. By employing techniques like distillation, reinforcement learning, and high-quality data curation, the Phi-4 models achieve a balance of size and performance, enabling efficient deployment in low-latency environments while maintaining robust reasoning abilities.
Phi-4-reasoning, a 14-billion parameter model, rivals significantly larger models in complex reasoning benchmarks, including mathematical problem-solving and Ph.D.-level science tasks. It utilizes supervised fine-tuning on curated reasoning demonstrations to produce detailed reasoning chains, achieving better performance than OpenAI o1-mini and DeepSeek-R1-Distill-Llama-70B. Phi-4-reasoning-plus builds upon this foundation, incorporating reinforcement learning and additional inference-time compute to further enhance accuracy.
Phi-4-mini-reasoning, a compact 3.8-billion parameter model, is tailored for environments with constrained computing or latency. It excels in step-by-step mathematical problem-solving across diverse difficulty levels, making it ideal for educational applications and lightweight deployment on edge devices. Despite its smaller size, it outperforms larger models like OpenThinker-7B and DeepSeek-R1-Distill-Qwen-7B in math benchmarks.
The Phi models are integrated across Windows devices, including Copilot+ PCs, leveraging NPU optimizations for efficient local deployment. They support applications like Outlook’s offline Copilot summary features and provide APIs for seamless integration into productivity tools. Developed with Microsoft’s responsible AI principles, the Phi family incorporates robust safety measures, including supervised fine-tuning, reinforcement learning from human feedback, and direct preference optimization, ensuring reliability and fairness in diverse use cases.
Available on Azure AI Foundry and HuggingFace, Phi-4 Reasoning models redefine the possibilities of small and efficient AI, delivering frontier-level reasoning capabilities in compact formats.
