MobileLLM-R1: Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes

Researchers developed MobileLLM-R1, a sub-billion parameter AI model that demonstrates strong reasoning capabilities using only 2T tokens of high-quality data instead of massive 10T+ token datasets. The 950M parameter model achieves superior performance on reasoning benchmarks compared to larger competitors while using only 11.7% of the training data compared to proprietary models like Qwen3.

• →MobileLLM-R1-950M achieves an AIME score of 15.5, vastly outperforming OLMo-2-1.48B (0.6) and SmolLM-2-1.7B (0.3) despite being smaller.
• →The research challenges the assumption that reasoning capabilities require massive datasets, showing 2T high-quality tokens are sufficient.
• →The model matches or surpasses Qwen3-0.6B performance while using only 11.7% of its proprietary 36T-token training corpus.
• →Complete training recipes, models, code, and data sources have been made publicly available for research.
• →This demonstrates efficient AI development through careful data curation rather than brute-force scaling approaches.