Models trade world knowledge for reasoning, study finds

GLM-5.2 scores 99.2% on AIME 2026 with ~40B active parameters, while Qwen3.5 9B fits in 6GB VRAM and doubles the next best sub-10B model. Yet SimpleQA leader Gemini 2.5 Pro still misses half of factual questions, and small models hallucinate 80-82% of the time.