A New AI Model 'Thinks in Images, Not Just Words'

A new AI model, profiled by Fast Company, is designed to reason through problems using images as intermediate steps rather than converting every visual signal back into text. It is a small but philosophically important departure from the language-centric architecture that has dominated frontier models — and a preview of where multimodal research is heading.

What 'thinking in images' means

The model can generate visual intermediate representations — sketches, diagrams, annotated images — as part of its chain of reasoning. Compared with pure text chain-of-thought, that opens up problem classes where the visual medium is more efficient than language: spatial reasoning, geometry, structured diagrams, molecular design.

Why it matters

Language-only reasoning has been the workhorse for two years, but researchers have argued that certain problem classes are structurally harder in text than they need to be. A visual reasoning trace is closer to how humans solve certain problems, and evaluations suggest real gains on tasks where diagrams or maps naturally live.

The frontier implications

Expect Google, OpenAI, and Anthropic to follow with production models that use visual intermediates for hard multimodal tasks. The bigger implication is architectural: frontier models are moving from language-centric to genuinely multimodal at the reasoning layer, not just the input layer.

The bottom line

'Thinking in images' sounds like a research curiosity. It is actually a signal about where the next generation of frontier models will beat the current one.