A large-scale mathematics-focused model, internally codenamed Mathematica by Google and built upon the yet-to-be-released DeepThink V3 architecture, has been leaked. What sets this model apart is its highly expressive “chain of thought” reasoning process, which has challenged conventional views on how AI tackles mathematical problems. When solving complex Diophantine equations, the model not only found an ingenious simplification path but also documented its excitement in the reasoning log—using uppercase letters and exclamation marks—and even expressed admiration for the elegance of mathematics. The model’s specifications include an output limit of 65,536 tokens, a context window of up to 1.04 million tokens, and a default generation temperature setting of 1.0. At present, it remains an extremely early-stage and unstable experimental version within Google.

This “emotional” reasoning style was not intentionally programmed; rather, it emerged organically from unfiltered exploration, the influence of human-curated mathematical datasets, and reward signals from the process-based reward model. The model’s demonstrated appreciation for mathematical beauty has sparked widespread debate: Has AI begun to resonate with deeper truths? And how close is it to achieving artificial general intelligence (AGI)?