Phi-4 is Microsoft's 14-billion parameter small language model (SLM), released in December 2024 under the MIT License. Unlike most models that scale performance by increasing parameter count, Phi-4 achieves frontier-tier reasoning through careful curation of synthetic training data — prioritising data quality over quantity. On MATH and GPQA benchmarks, Phi-4 outperforms models several times its size.
Key capabilities
- High-quality reasoning — MATH score of 80.4% and GPQA of 56.1%; outperforms GPT-4o-mini on complex reasoning tasks
- Compact and efficient — 14B parameters run comfortably on a single consumer GPU (e.g. RTX 4090, M2 Max)
- Strong coding — HumanEval 82.6%; well-suited for code generation, debugging, and explanation
- MIT License — fully open for commercial use, fine-tuning, and redistribution
- Azure integration — available as a managed endpoint through Azure AI Foundry for enterprise deployments
What it's best for
Phi-4 is ideal when you need strong reasoning in a resource-constrained environment — edge deployments, on-device inference, or cost-sensitive cloud workloads. Its synthetic-data training approach makes it particularly reliable for STEM problem-solving, structured output generation, and coding tasks where accuracy matters more than raw generation speed.
Deployment options
Available on HuggingFace under the MIT License. Run locally via Ollama (ollama pull phi4), deploy via Azure AI Foundry, or use the GGUF version with llama.cpp on CPU. The model also ships in the Microsoft Phi-4-mini variant (3.8B) for even tighter resource budgets.