3.35B Open Model Achieves 93% In-Language Reasoning Across 60 Languages via Data Mixing Alone

Most reasoning models default to English even when prompted in other languages, locking out billions of non-English speakers. This paper shows that data mixing and scheduling in supervised fine-tuning alone—no new architecture—transfers reasoning generalization across 60 typologically diverse languages at 93%+ rate, and the weights are released openly.