Microsoft open-sources small language model Phi-4 in Hugging Face

Microsoft announced that it has released its small language model Phi-4 as completely open source. Thus, the model became available for download on the Hugging Face platform.

Phi-4 is the fourth version of the series of small language models that Microsoft introduced in 2023. With 14 billion parameters, the model outperforms its larger rivals in areas such as mathematical reasoning, multi-tasking language understanding and functional code generation. For example; It was stated that Google passed larger models such as Gemini Pro and GPT-4o-mini with over 80 percent success in the challenging MATH and MGSM benchmark tests.

The model is based on a dense and decoder-only transformer architecture. This approach provides a more efficient structure by reducing computational costs. Phi-4 was trained on high-quality datasets of math, coding, and logic-based data over 9.8 trillion tokens. It also offers limited support for non-English apps, using 8 percent multilingual content.

Phi-4 was first available under a research license on Microsoft’s Azure AI Foundry platform in December 2024. However, this access was limited. Now, the model has been made publicly available under the MIT license via Hugging Face. This license allows the model to be used freely in commercial projects.

Sharing the model’s weights allows other researchers and developers to customize the model and adapt it to different projects. With this move, Microsoft aims to support transparency and innovation in the field of artificial intelligence.