Microsoft Releases Phi-4-Reasoning-Vision-15B: A Compact Multimodal Model for Math, Science, and GUI Understanding

2026-03-06 · Source: Machine Learning ML & Generative AI News · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, quick

Summary

Microsoft has released Phi-4-reasoning-vision-15B, a 15B open-weight multimodal reasoning model. This model integrates Phi-4-Reasoning with SigLIP-2 using a mid-fusion architecture, enabling it to process image-and-text tasks with reduced computational demands compared to larger vision-language models. Trained on 200B multimodal tokens, its design prioritizes preserving high-resolution visual detail for dense documents and interfaces, alongside a mixed reasoning setup that allows for dynamic switching between direct responses and explicit reasoning. Phi-4-reasoning-vision-15B is optimized for applications in math, science, document understanding, OCR, and GUI grounding, demonstrating strong performance on benchmarks including AI2DTEST, ChartQATEST, MathVistaMINI, OCRBench, and ScreenSpotv2.

Key takeaway

For AI Scientists and Machine Learning Engineers developing multimodal applications, Phi-4-reasoning-vision-15B presents a compelling option for achieving strong performance with lower compute. Its specialized design for high-resolution visual detail and mixed reasoning makes it particularly suitable for tasks involving complex documents or graphical user interfaces. Consider evaluating this 15B model for your next project requiring efficient vision-language integration.

Key insights

Phi-4-reasoning-vision-15B offers compact, multimodal reasoning for diverse tasks with lower compute.

Principles

• Preserve high-resolution visual detail.
• Enable mixed reasoning capabilities.

Method

Combines Phi-4-Reasoning with SigLIP-2 in a mid-fusion architecture, trained on 200B multimodal tokens for image-and-text tasks.

In practice

• Apply to math and science problems.
• Use for document understanding and OCR.
• Integrate for GUI grounding tasks.

Topics

• Multimodal Models
• Vision-Language Models
• Mid-fusion Architecture
• Document Understanding
• GUI Grounding

Code references

• microsoft/Phi-4-reasoning-vision-15B

Best for: Machine Learning Engineer, Computer Vision Engineer, AI Scientist, AI Researcher, AI Engineer, AI Student

Related on AIssential

• Phi-4-reasoning-vision and the lessons of training a multimodal reasoning model — Efficient multimodal reasoning models can achieve high performance with less data and compute through careful design and data curation.
• Imaginative Perception Tokens Enhance Spatial Reasoning in Multimodal Language Models — Imaginative Perception Tokens (IPT) improve VLM spatial reasoning by externalizing unobserved perceptual information.
• OmniSch: A Multimodal PCB Schematic Benchmark For Structured Diagram Visual Reasoning — Current LMMs struggle with fine-grained visual grounding and topological reasoning for complex PCB schematics, necessitating specialized benchmarks.
• Visual-OPSD: Cross-Modal On-Policy Self-Distillation for Efficient Unified Multimodal Reasoning — The generation pathway of visual thoughts encodes reasoning that can be distilled for efficient, text-only multimodal models.
• ReasonCLIP-58M: Visually Grounded Commonsense Reasoning Supervision for CLIP — Integrating structured reasoning supervision into CLIP-style models significantly enhances their expressive capacity for visually grounded commonsense and compositional reasoning.
• Granite 4.0 3B Vision: Compact Multimodal Intelligence for Enterprise Documents — Granite 4.0 3B Vision offers compact, modular multimodal intelligence for enterprise document understanding.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning ML & Generative AI News.