AI Research Atlas

Phi-4-reasoning-vision-15B

Microsoft · 4 March 2026

15B open vision-language model that decides when to reason; trained on only 200B multimodal tokens, 75.2% on MathVista and 88.2% on ScreenSpot v2.

Mid-fusion of a SigLIP-2 encoder with the Phi-4-reasoning backbone, trained on a mix of ~20% reasoning and 80% direct-answer data so perception tasks skip long chains of thought.

Date
Wednesday, 4 March 2026
Lab
Microsoft
Kind
open-weights
Access
open weights

Figures

MeasureValueMeasured by
MathVista_MINI / ScreenSpot_v275.2% / 88.2%
MMMU val 54.3%; 200B multimodal tokens		company

Sources

  1. www.microsoft.com/en-us/research/blog/phi-4-reasoning-vision-and-the-lessons-of-training-a

This record was checked against its sources on 6 October 2026. How we check

Related