Microsoft's New 15B AI Model Matches Larger Rivals Using One-Fifth the Training Data

Microsoft's new 15-billion-parameter AI model, Phi-4-reasoning-vision-15B, matches and even outperforms much larger rivals while using only one-fifth the training data, thanks to a 'mixed reasoning' design that intelligently switches between deep analytical thinking and fast direct responses depending on task complexity.

• Microsoft releases Phi-4-reasoning-vision-15B, a 15-billion-parameter open-weight multimodal AI model available on Microsoft Foundry, HuggingFace, and GitHub, capable of processing images and text to solve complex math, science, and visual tasks while matching or outperforming much larger models.
• The model is trained on roughly 200 billion tokens — about one-fifth the data used by rival models — thanks to meticulous data curation, manual quality review, and response regeneration using GPT-4o, challenging assumptions about the necessity of massive training datasets.
• A key innovation is the model's 'mixed reasoning' design, where only 20% of training samples include chain-of-thought reasoning, allowing it to apply deep thinking for math and science while defaulting to fast, direct responses for simpler visual tasks like captioning, optimizing both accuracy and efficiency.