On Wednesday, Microsoft announced the open-sourcing of its Phi-4 small language model. This new addition to the Phi series was initially unveiled last month, but access was restricted to the Azure AI Foundry at that time. The tech giant from Redmond had previously indicated its intent to make the source code publicly available, and now, users can find the reasoning-focused AI model on Hugging Face. The model is available for both academic and commercial purposes.
Microsoft Open-Sources Phi-4 AI Model
Shital Shah, a technical staff member at Microsoft AI, announced the model’s availability on X (formerly Twitter), indicating that the Phi-4 model’s weights can now be accessed through Hugging Face. The model is licensed under the MIT license, allowing for academic and commercial use. Those interested can view the model details here.
The Phi-4 model, released eight months after its predecessor Phi-3, is designed to significantly enhance performance in resolving complex reasoning tasks, particularly in mathematics. With a context window of 16,000 tokens, it has been trained on a massive dataset totaling 9.8 trillion tokens.
The Hugging Face listing outlines the diverse sources of the training data, which include high-quality publicly available educational materials, synthetic data covering various subjects, acquired academic texts, and question-and-answer datasets, alongside data formatted for chat supervision.
Importantly, Phi-4 is a text-only model, capable of processing only text for both input and output. It features 14 billion parameters and utilizes a dense decoder-only Transformer architecture, as per Microsoft’s specifications.
Upon its launch, Microsoft shared benchmark evaluations of the model, claiming that Phi-4 surpasses the Gemini 1.5 Pro model in solving math competition problems.
In addition to being available via Hugging Face, the Phi-4 AI model can also be accessed through Microsoft’s Azure AI Foundry. This platform assists developers and enterprises in managing AI-related risks while offering features such as prompt shields, groundedness detection, and content filters. These safety measures can be integrated into applications via the company’s application programming interface (API).
