A new language model, VibeThinker-3B, with just three billion parameters, is competitive with much larger models on math and coding tasks, but falls behind on tasks requiring broad factual knowledge. This suggests that logical reasoning can be compressed into smaller models, but factual knowledge still requires larger models. The model’s performance is due to multi-stage post-training applied to a base model. Why it matters: this research has significant implications for the development of more efficient AI models, highlighting the importance of understanding how different types of knowledge are structured and compressed.