Why 2026's AI Labs Are Building Models That Know Less but Reason Better

A growing trend in AI development sees leading laboratories deliberately reducing the factual knowledge stored in large language models in favor of stronger reasoning capabilities. The shift is driven by the insight that storing facts in model weights is costly, quickly outdated, and does not generalize well to novel situations. A striking example emerged when Alibaba's Qwen 3.8 27B model spent 21 minutes and over 22,000 reasoning tokens to respond to a simple prompt asking it to draw an SVG circle, autonomously expanding the task to explore its artistic potential. Complementary advances such as Multi-Token Prediction are delivering throughput gains of around 72%, helping offset the speed cost of deeper reasoning. The emerging architecture places factual knowledge in external retrieval systems while reserving model parameters almost entirely for logical and compositional reasoning.