Phi-4 Reasoning Vision 15B: Multimodal AI with Chain‑of‑Thought Power

Phi-4-Reasoning-Vision-15B is an open‑weight multimodal model released by Microsoft on March 4, 2026. It combines the Phi‑4‑Reasoning language backbone (5 B–15 B parameters) with a SigLIP‑2 vision encoder in a mid‑fusion architecture, allowing it to ingest both text and images and generate text responses. The model supports the `image-text-to-text` pipeline tag and is built for English‑only vision‑language tasks. Training used 240 NVIDIA B200 GPUs for four days, integrating supervised fine‑tuning on a curated mix of reasoning and perception data, and includes a safety post‑training step that teaches refusal behavior for harmful content.

The model excels at visual question answering and reasoning over complex visual inputs. Benchmark results show accuracy of 84.8 % on AI2D (diagram understanding), 83.3 % on ChartQA, 75.2 % on MathVista‑MINI, 88.2 % on ScreenSpot‑V2 (GUI grounding), and 76 % on OCRBench. It can switch between a fast "NOTHINK" mode for straightforward perception tasks and a structured "THINK" mode that produces chain‑of‑thought reasoning using `<think>...</think>` blocks, making it suitable for both captioning and multi‑step problem solving.

Phi-4‑Reasoning‑Vision‑15B is positioned for resource‑constrained environments while still delivering high‑quality multimodal reasoning. Its primary use cases include scientific and mathematical problem solving from images (handwritten equations, charts, diagrams), computer‑use agents that interpret screen content and locate GUI elements, as well as general tasks such as image captioning, OCR, and object grounding. The model is released under an MIT license, distributed via Hugging Face, GitHub, and Azure AI Foundry, and requires PyTorch ≥ 2.7.1, Transformers ≥ 4.57.1, and optionally vLLM for inference.