QED-Nano: Teaching a Tiny Model to Prove Hard Theorems

Researchers developed QED-Nano, a 4B parameter AI model that achieves competitive performance on Olympiad-level mathematical proofs despite being much smaller than proprietary systems. The model uses a three-stage training approach including supervised fine-tuning, reinforcement learning, and reasoning cache expansion to match larger models at a fraction of the inference cost.

• →QED-Nano demonstrates that small open-source models can compete with large proprietary AI systems on complex mathematical reasoning tasks.
• →The 4B parameter model outperforms much larger open models like Nomos-1 and GPT-OSS-120B while approaching Gemini 3 Pro's performance.
• →The three-stage training pipeline includes supervised fine-tuning, reinforcement learning with rubric-based rewards, and reasoning cache expansion.
• →The complete training pipeline, models, datasets, and evaluation code have been released as open-source resources.
• →This breakthrough could significantly reduce the computational costs and barriers to accessing high-performance mathematical reasoning AI.