Microsoft releases Phi-4-mini and Phi-4-multimodal

Background

Phi-4-mini and Phi-4-multimodal were announced as the newest small language models in Microsoft's Phi family.^[1][2]

What happened

On 26 February 2025 Microsoft announced Phi-4-mini and Phi-4-multimodal as the newest Phi small language models.^[1][2] Phi-4-mini is a 3.8-billion-parameter text model. Phi-4-multimodal has 5.6 billion parameters and handles text, vision and speech.^[2] Both were released under the MIT licence on Hugging Face.^[2]

The two models became available the same day in Azure AI Foundry and GitHub Models.^[1][3]

Key details

Phi-4-mini supports a 128K-token context and a 200K-token vocabulary.^[4] A context of this size means the model can take in a long document or conversation at once, and a larger vocabulary lets it represent text with fewer tokens.

Phi-4-mini was trained on 5 trillion tokens. It scores 67.3 on MMLU, a multi-subject knowledge and reasoning test, and 88.6 on GSM8K, a set of grade-school maths word problems.^[4]

Impact

The release gave developers two small models under one licence: a text-only model and a model that also processes images and speech. Both were offered on Azure AI Foundry and GitHub Models on the day of release, in addition to Hugging Face.^[1][2][3][4]

Terms used here

A token is a chunk of text, and a vocabulary is the set of tokens a model uses. A larger vocabulary means the model can represent the same text with fewer tokens. A multimodal model accepts more than one kind of input, here text, images and speech or audio. Azure AI Foundry is Microsoft's platform for building and running AI models, and GitHub Models is a GitHub service for trying models. Hugging Face is the online hub where the weights were published, and the MIT licence allows others to use, modify and redistribute them.

Azure AI Foundry and GitHub Models are both places where developers can call a model without first setting up their own servers.