Microsoft releases Phi-4-reasoning-vision-15B

Background

Phi-4-reasoning-vision-15B is a vision-language model, a type of model that reads both images and text. It was built on the reasoning backbone of Phi-4-reasoning, which Microsoft had already released as a text model, and it adds a SigLIP-2 vision encoder to read images.^[1][2]

What happened

On 4 March 2026 Microsoft released Phi-4-reasoning-vision-15B as an open-weight vision-language model.^[1][2] The model combines the Phi-4-reasoning backbone with a SigLIP-2 vision encoder.^[1] It chooses whether to reason step by step or answer directly, depending on the task.^[1][2]

Key details

The model was trained on 240 NVIDIA B200 GPUs for 4 days.^[1] On the ScreenSpot v2 test, which measures locating elements on a screen, it scored 88.2. On MathVista-mini, a visual maths test, it scored 75.2.^[1]

Its weights are released under the MIT licence, which lets others use, modify and redistribute the model.^[1]

Impact

The release added a vision model with a built-in choice between reasoning and direct answers to the open-weight models Microsoft publishes. The model card reports scores on screen-location and visual maths tests.^[1][2]

Terms used here

A vision-language model accepts images and text as input and answers in text. A vision encoder is the component that converts an image into a sequence of numerical tokens that the language model can process. An open-weight model publishes its trained parameters so that others can download and run them, subject to the licence. ScreenSpot tests whether a model can locate elements on a screen, such as buttons or fields in an interface. MathVista tests mathematical reasoning over images such as charts and diagrams. Computer-use agents operate a computer by looking at the screen and choosing where to click or type, so accurately locating screen elements is a core skill for them. The MIT licence allows others to use, modify and redistribute the model.