Phi-4-reasoning-vision-15B

Claims and evidence

• Released 4 March 2026 ● Primary source[1]Published date; first paragraph [2]header: Release Date [3]submission history, v1

• Status AvailableOpen weights still downloadable from the microsoft organisation on Hugging Face (repository not gated, weight files listed, checked through the Hugging Face API). ● Primary source[2]header: License MIT; 2.3 Distribution Channels

• Derived from (other) Phi-4-reasoningMid-fusion model: a vision encoder feeds visual tokens into the Phi-4-reasoning language backbone, followed by supervised fine-tuning. The Hugging Face repository metadata labels it a fine-tune of Phi-4-reasoning. ● Primary source[2]header: Model Dependencies; 1. Model Overview [1]Model architecture: Vision encoder and image processing

• Change · Modality Adds image input: a SigLIP-2 vision encoder turns images into tokens that are projected into the Phi-4-reasoning language backbone (mid-fusion).Compared with Phi-4-reasoning ● Primary source[2]1. Model Overview
• Change · Reasoning Trained on a mix of reasoning and non-reasoning data, so one model can reason step by step or answer directly, selected with think and nothink mode tokens.Compared with Phi-4-reasoning ● Primary source[2]1. Model Overview; 2.4 Input Formats
• Change · Other Microsoft positions it for computer-use agent tasks, such as reading screens and locating interface elements.Compared with Phi-4-reasoning ● Primary source[1]first paragraph [2]2.1 Primary Use Cases

• Input text, image ● Primary source[2]header: Inputs
• Output text ● Primary source[2]header: Outputs
• Feature Reasoning modeOne model switches between chain-of-thought answers (think) and direct answers (nothink) via mode tokens. ● Primary source[2]1. Model Overview; 2.4 Input Formats [3]abstract
• Feature Computer usePositioned for computer-use agent tasks: interpreting screens and localising GUI elements. ● Primary source[2]2.1 Primary Use Cases [1]first paragraph
• Open weights YesMIT licence. ● Primary source[2]header: License MIT [1]first paragraph
• Context window 16,384 tokens tokens ● Primary source[2]header: Context Length
• Parameters 15B ● Primary source[2]header: Parameters [1]first paragraph
• Access Open-weights download, APIapi stands for Microsoft Foundry, Microsoft's own model platform; code is also on GitHub. ● Primary source[1]first paragraph [2]2.3 Distribution Channels

Lineage

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Microsoft on AI Radar →