Phi-4-reasoning-vision

Developer: Microsoft Corporation Authorized Representative: Microsoft Ireland Operations Limited, 70 Sir John Rogerson's Quay, Dublin 2, D02 R296, Ireland Release Date: March 4, 2026 License: MIT Parameters: 15B Context Length: 16,384 tokens Inputs: Text and Images Outputs: Text Training GPUs: 240 B200s Training Time: 4 days Training Dates: February 3, 2025 – February 21, 2026 Model Dependencies: Phi-4-Reasoning

1. Model Overview

Phi-4-Reasoning-Vision-15B is a compact open-weight multimodal reasoning model built on the Phi-4-Reasoning language model backbone and the SigLIP-2 vision encoder, using a mid-fusion architecture. In this architecture, the vision encoder first converts images into visual tokens, which are then projected into the language model's embedding space and injected into the pretrained language model. This approach leverages the strengths of both pretrained components while keeping training and inference costs manageable. The model employs a dynamic resolution vision encoder with up to 3,600 visual tokens, enabling high-resolution image understanding critical for tasks such as GUI grounding and fine-grained document analysis. Bidirectional attention is applied within images (intra-image) to improve spatial reasoning without the overfitting risks observed with broader bidirectional schemes.

Phi-4-Reasoning-Vision-15B is trained with Supervised Fine-Tuning (SFT) on a carefully curated mixture of reasoning and non-reasoning data. Rather than training separate models for each mode, the model operates as a single system that can invoke extended chain-of-thought reasoning (using ... blocks) for tasks like mathematical and scientific reasoning, or default to direct inference (tagged with ``) for perception-focused tasks such as captioning, object detection, and grounding. The training data consists primarily of meticulously filtered and improved open-source vision-language datasets, supplemented by high-quality domain-specific data from internal Microsoft teams and targeted data acquisitions. This data-centric approach, combined with moderate training compute requirements (240 NVIDIA B200 GPUs for 4 days), distinguishes Phi-4-Reasoning-Vision-15B from models that rely on substantially more training data and compute.

1.1 Alignment Approach

Phi-4-Reasoning-Vision-15B has adopted a safety post-training approach leveraging a combination of open-source and in-house generated synthetic datasets. The safety alignment is achieved through Supervised Fine-Tuning (SFT) using data that includes both helpfulness and harmlessness examples, as well as targeted questions and answers across multiple safety categories. The model's training data explicitly includes safety-oriented samples designed to teach appropriate refusal behavior for harmful content categories including hate speech, violence, self-harm content, and sexually explicit material. Automated red teaming was performed on Azure to assess safety risks including groundedness, jailbreak susceptibility, harmful content generation, and copyright violations for protected material.

2. Usage

2.1 Primary Use Cases

Phi-4-Reasoning-Vision-15B is designed for general-purpose multimodal AI systems and applications that require vision-language understanding with selective reasoning capabilities, particularly in memory- or compute-constrained environments. The model excels in two primary domains:

• Scientific and mathematical reasoning over visual inputs: such as solving math problems presented as handwritten equations or diagrams, extracting and reasoning over quantitative information in documents, charts, and tables, and supporting multi-step reasoning in educational or scientific analysis contexts.
• Computer-use agent (CUA) tasks: such as interpreting screen content, localizing interactive GUI elements, and selecting actions within graphical user interfaces.

The model is also capable of general multimodal tasks including image captioning, visual question answering, optical character recognition, object localization, and grounding. Its hybrid reasoning design allows it to produce fast, direct responses for perception-focused tasks while engaging in structured chain-of-thought reasoning when the task benefits from it, making it suitable as a building block for generative AI-powered features across a range of applications.

2.2 Out-of-Scope Use Cases

Phi-4-Reasoning-Vision-15B is not specifically designed or evaluated for all downstream purposes. Developers should consider common limitations of vision-language models as they select use cases, and evaluate and mitigate for accuracy, safety, and fairness before using within a specific downstream use case, particularly for high-risk scenarios.

The model is trained primarily on English text and image-text pairs. Languages other than English may experience degraded performance. The model should not be used in scenarios that could have consequential impact on legal status or the allocation of resources or life opportunities (e.g., housing, employment, credit) without further assessments and additional debiasing techniques. It is not suitable for providing medical diagnoses, le