Models Are Getting Dumber on Purpose

Models are being intentionally designed to prioritize reasoning abilities over factual knowledge. While powerful reasoning scores have been achieved in smaller models, these models struggle with basic factual recall. For example, Qwen3.5 4B and 9B achieve hallucination rates of 80 to 82% on knowledge benchmarks. To achieve a model with frontier-quality reasoning and a compact size, facts are removed and replaced with runtime retrieval from knowledge bases, tool calls, and web search.

This allows for a 20 to 40B model at 4-bit quantization to fit on a single consumer GPU, without needing to store massive amounts of outdated facts. The tradeoff enables more efficient models that can keep up with the constantly changing facts of the world, without the heavy computational cost of maintaining a large knowledge base within the model itself.