Microsoft fully open-sources Phi-4 on Hugging Face with an MIT license

Phi-4 is Microsoft’s 14-billion-parameter open model built for heavy reasoning and logic, and its official weights are now fully accessible on Hugging Face under a permissive MIT license.

While Microsoft first introduced Phi-4 in December on its Azure AI Foundry platform under a research license, developers have been clamoring for raw weights. After unofficial bootlegs started popping up on Hugging Face, Microsoft decided to make it official and completely unrestricted for commercial use.

Big reasoning in a 14B package

Despite its compact 14B parameter footprint, Phi-4 is designed to punch well above its weight class in analytical domains.

On challenging math benchmarks like MATH and MGSM, the model scores above 80%, outpacing significantly larger closed systems including Google’s Gemini Pro and OpenAI’s GPT-4o-mini. It also posts competitive scores on HumanEval, making it a capable engine for automated code generation and technical problem-solving.

Specs, data recipe, and how to run it

Under the hood, Phi-4 is a dense decoder-only Transformer with a 16K context window.

Its intelligence comes down to synthetic data engineering. Microsoft trained the model across 1,920 H100 GPUs for 21 days, consuming 9.8 trillion tokens. The dataset combines filtered public documents, synthetic textbook-style materials focusing on math and logic, and academic Q&A datasets, with multilingual text making up roughly 8%. It was further aligned using supervised fine-tuning and direct preference optimization.

Running it locally is straightforward. You can pull the weights directly via Hugging Face Transformers, serve an OpenAI-compatible endpoint using vLLM or SGLang, or fire it up using Docker Model Runner without requesting any special API access.

Why it matters

For independent developers and engineering teams, 14B represents a hardware sweet spot: it is small enough to host locally or at the edge on accessible hardware, yet powerful enough to tackle actual STEM workflows.

We think releasing this under an MIT license is a major win for builders who want top-tier reasoning without cloud vendor lock-in. If you need a reliable, commercially usable foundation model for coding or complex math that runs privately on your own infrastructure, Phi-4 is one of the best current options to test.