The dominant trend in artificial intelligence of late has been clear: bigger models, more data, more compute. But the cost of training and running those systems keeps rising as companies pour enormous computing resources into ever larger models. It’s this tension that has led researchers at Microsoft to explore a different path: training more compact models on carefully curated reasoning data to see how far they can go.
In a blog post last week, Microsoft described Phi-4-Reasoning-Vision-15B, a multimodal model designed to handle reasoning…
