Phi is a family of small language models (SLMs) from Microsoft. Developers use them to add generative AI to applications, either in the cloud or running directly on a device without cloud connectivity. The page describes the models as lightweight and resource efficient, with fast response times for real-time applications.
The family includes several series, from Phi-1 through Phi-4, with newer models adding multilingual support, function calling and multimodal inputs for text, audio and vision. Developers can fine-tune the models with their own domain-specific data to improve results for a particular task.
Phi is aimed at developers and organisations that want a capable model for tasks with tight cost or latency limits, or where the model needs to run locally for more control over its environment. The page says the models are available as open source under the MIT License, through Microsoft Foundry, Hugging Face and Ollama.
On cost, the page says Phi models can be used for free for real-time deployment through Microsoft Foundry or Hugging Face. It also offers pay-as-you-go billing through model as a service (MaaS) with inference APIs in Azure. No specific dollar prices are listed on the page.
Microsoft says the models are developed in line with its AI principles of accountability, transparency, fairness, reliability and safety, privacy and security, and inclusiveness. The newer models are described as designed for production environments.
