Microsoft Research · 2026-04-10 · notable
Phi-4-Reasoning-Vision-15B — Microsoft's Compact Multimodal Reasoner: 88.2% ScreenSpot v2, Open-Weight
Phi-4-Reasoning-Vision-15B combines SigLIP-2 vision with the Phi-4-Reasoning language backbone via mid-fusion. At 15B parameters it achieves 88.2% ScreenSpot v2, 75.2% MathVista, and 83.3% ChartQA — outperforming larger models while staying open-weight under a permissive license.
Microsoft's 15B open-weight model combines visual reasoning with math and GUI understanding — and knows when NOT to think.
Key specs
| Parameters | 15B |
|---|---|
| Screenspot v2 | 88.2% |
| Mathvista | 75.2% |
| Chartqa | 83.3% |
| Mmmu | 54.3% |
What is it?
Phi-4-Reasoning-Vision-15B is Microsoft's compact multimodal reasoning model, released open-weight on March 4, 2026 via Microsoft Foundry and Hugging Face. It extends the Phi-4-Reasoning language backbone with a SigLIP-2 vision encoder using a mid-fusion architecture. A blog post with detailed training insights was published in April 2026.
How does it work?
The model uses a mid-fusion approach combining SigLIP-2's vision encoder with the Phi-4-Reasoning backbone, processing high-resolution images via dynamic resolution up to 3,600 visual tokens. Training data was split roughly 20% reasoning-focused (with explicit chain-of-thought traces) and 80% perception-focused (direct response), enabling the model to invoke structured reasoning for math and science while defaulting to fast response for perception tasks.
Why does it matter?
GUI grounding benchmarks like ScreenSpot v2 reflect real agentic workflows: reading screens, clicking buttons, understanding interfaces. At 88.2% on ScreenSpot v2 and comparable to larger models at far lower compute, Phi-4-Reasoning-Vision opens the door for capable multimodal agents on edge hardware and laptops.
Who is it for?
Developers building computer-use agents, document understanding pipelines, or multimodal scientific tools who need an open-weight model with strong reasoning at low deployment cost.