Models Are Getting Dumber on Purpose

Researchers are intentionally designing artificial intelligence models to prioritize reasoning abilities over factual knowledge. While larger models boast higher reasoning scores, their factual recall remains limited. For instance, GLM-5.2 scores 99.2% on AIME 2026 with 40 billion parameters active per token, while Qwen3.5 achieves 91.3% with just 17 billion active parameters.

Even smaller models like Qwen3.5 9B, which fits in 6GB of VRAM, demonstrate impressive reasoning capabilities yet struggle with factual accuracy when confronted with plain questions. The reasoning process in these models compresses information more efficiently than facts, allowing them to learn general problem-solving techniques well.

However, facts require more storage space and cannot be retained indefinitely due to the rapidly evolving nature of information. As a result, researchers are sacrificing factual content in favor of enhanced reasoning abilities, leading to models that are adept at problem-solving but lack deep knowledge in specific domains. This trade-off enables the creation of smaller, more lightweight models that can run efficiently on consumer-grade hardware, while still providing valuable reasoning capabilities.

These models rely on external knowledge bases, tool calls, or web searches to retrieve the necessary information when needed, rather than relying on their internal storage. The result is a model that focuses on reasoning prowess and adapts to changing circumstances by leveraging external resources, making it a more flexible and cost-effective solution in the long run.