Phi-4 Reasoning Vision 15B
Phi-4 Reasoning Vision 15B combines the Phi-4 Reasoning language backbone with a SigLIP-2 image encoder. It can select a direct response or an extended reasoning mode for visual questions.
What distinguishes this release
The release uses a mid-fusion design and a dynamic-resolution encoder with up to 3,600 visual tokens. Its card covers scientific diagrams, document analysis and screen-element grounding, and identifies English as its supported language focus.Phi-4 Reasoning Vision 15B official model card
Build the task around its input and output
Practical guidance. For screen tasks, treat returned locations as predictions that must match the current screenshot. Keep the image dimensions and coordinate convention with the result. Recheck the screen before a consequential action, because a correct location in an old image can be wrong now.
Review the result in its intended setting
Practical guidance. For a chart or diagram, ask the model to identify the visible values before drawing a conclusion. Review perception errors separately from calculation errors. This helps determine whether to improve the image input or change the reasoning step.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | microsoft/Phi-4-reasoning-vision-15B |
| Context window | 16,384 tokens Published checkpoint context; serving engines may impose lower limits Phi-4 Reasoning Vision 15B official model card |
| Model size | 15B multimodalPhi-4 Reasoning Vision 15B official model card |
| Input and output | Text and images in; text outPhi-4 Reasoning Vision 15B official model card |
| Weight license | MITPhi-4 Reasoning Vision 15B official model card |