The quest for artificial general intelligence (AGI) hinges significantly on developing reasoning models that can autonomously learn, adapt, and evolve, much like human cognition. Current large language models (LLMs) exhibit impressive capabilities in language understanding and generation, but often fall short in true reasoning and problem-solving, especially in open-ended environments. Existing self-play methodologies have shown promise in specific domains, yet struggle with generalization, relying heavily on predefined reward models or fixed functionalities. To address these limitations, a novel paradigm, "Absolute Zero," is proposed, aiming to redefine the very essence of reasoning model training. This paradigm focuses on enabling the model to simultaneously define tasks that maximize learnability and to solve them effectively, thus fostering self-evolution through self-play without external data reliance.