The future of large language models may not lie in sheer size, but in smarter design. For years, the AI community has operated under the assumption that bigger models—those with more parameters and massive training datasets—would naturally lead to better reasoning and intelligence. This belief, rooted in the idea that general intelligence emerges from scale, has driven the development of models with billions of parameters and trillions of training tokens. However, growing evidence suggests this approach has serious limitations. Many of today’s most powerful models are not truly reasoning. Instead, they are highly sophisticated imitators, relying on patterns from their training data to generate plausible-sounding answers. They often resort to verbose, step-by-step "thinking out loud" through chain-of-thought prompting, producing large numbers of tokens that are inefficient and frequently irrelevant. Worse, they lack the flexibility to adapt their thinking process based on problem difficulty. They cannot pause, reflect deeply, or allocate more mental effort to complex parts of a problem—features central to human cognition. Enter Hierarchical Reasoning Models (HRMs), a new architecture proposed by Wang et al. (2025). HRMs represent a fundamental shift in how models process information. Rather than forcing models to verbalize their thoughts, HRMs allow them to reason silently and fluidly within a high-dimensional latent space—closer to how humans intuitively solve problems before forming words. At its core, HRM is a two-tiered system: a slow, strategic High-level (H) module that sets the overall plan, and a fast, execution-focused Low-level (L) module that carries out the plan through iterative computation. The H-module defines a strategy, such as “explore paths moving downward and right” in a maze. The L-module then explores possible routes, backtracking when needed, and returns its findings to the H-module. Based on this feedback, the H-module adjusts its strategy—perhaps shifting to “now explore rightward paths”—and the cycle repeats. A key innovation is Adaptive Computation Time (ACT), which allows the model to decide when to stop thinking. After each full cycle of reasoning, a small network evaluates confidence in the current answer. Using a Q-learning framework, the model learns whether to halt and output a result or continue pondering. This enables intelligent efficiency: easy problems are solved quickly, while difficult ones receive more processing time—without any fixed number of steps. HRMs have demonstrated impressive results. On 30×30 maze and Sudoku challenges, they outperform all major chain-of-thought models, which often fail entirely. Even more striking, the HRM achieves this with just 27 million parameters, trained from scratch on only about 1,000 examples per task—far less than the massive datasets used by industry giants. It requires no expensive pre-training or complex prompt engineering. On the Abstraction and Reasoning Corpus (ARC-AGI), a benchmark for fluid intelligence, the HRM scores 40.3%—surpassing much larger models like o3-mini (34.5%) and Claude 3.7 (21.2%), despite its small size. Importantly, HRM’s performance scales almost linearly with compute, unlike standard transformers, which see diminishing returns. The model can leverage extra computation meaningfully, showing true depth of reasoning. The ACT mechanism further highlights its efficiency. While a fixed-step model uses the same number of steps for every problem, the HRM with ACT averages just 1.5 steps for easy tasks, achieving peak accuracy with far less resource use. These findings challenge the long-held belief that scale equals intelligence. Instead, they suggest that architectural innovation—particularly hierarchical, adaptive reasoning—may be the true path forward. The next generation of language models may not be large, but they could be far smarter.