Big reasoning power, small models
Small open-weight models (3.8B/14B) delivering powerful reasoning for math/science/code, rivaling larger LLMs. Available on Azure AI Foundry & HF.
Phi-4 Reasoning models are optimized for Phi Silica (NPU-optimized variant) and will run on Copilot+ PC NPUs using ONNX optimization. This enables blazing-fast time-to-first-token responses and power-efficient token throughput, allowing concurrent invocation with other applications while providing offline reasoning capabilities.
Phi-4 Reasoning models are positioned as the backbone of agentic applications. They balance size and performance through distillation and reinforcement learning, enabling resource-limited devices to perform complex reasoning tasks efficiently. Phi-4-reasoning-plus uses 1.5x more inference-time tokens for higher accuracy, making it suitable for sophisticated multi-step agent workflows.
Phi-4 Reasoning is a 14B parameter open-weight model trained via supervised fine-tuning on high-quality reasoning demonstrations. It generates detailed reasoning chains leveraging inference-time compute, achieving performance comparable to much larger models (DeepSeek-R1 at 671B parameters) on AIME 2025 and other mathematical benchmarks, while remaining small enough for low-latency environments and edge deployment.
Phi-4-mini-reasoning is a 3.8B parameter model optimized for mathematical reasoning and step-by-step problem solving. Trained on over one million diverse math problems from middle school to Ph.D. level, it outperforms models twice its size on math benchmarks and is ideal for educational applications, embedded tutoring, and lightweight edge/mobile deployment.
Phi-4 Reasoning models are available as open-weight models on both Azure AI Foundry and HuggingFace, enabling users to tinker, tweak, and customize within projects. This allows organizations to fine-tune the models on proprietary data, integrate them into custom pipelines, and maintain full control over deployment without vendor lock-in.