MiMo

MiMo is a 7 billion-parameter language model series created by Xiaomi with expertise in mathematical and coding reasoning abilities. It achieves similar performance to larger models thanks to unique pre-training and post-training techniques.

Visit Website

Prediction of Multiple TokensImproved training objectives aimed at enhancing model accuracy and speeding up inference times
Enhanced Pre-Training WorkflowUtilizes multidimensional data filtering and synthetic reasoning data generation techniques to enhance the density of reasoning patterns.
Highly Advanced Reinforcement Learning Training FrameworkFeatures a seamless rollout engine that delivers 2.29x faster training and 1.96x faster validation by continuously rolling out updates and asynchronously computing rewards.
Reward for Writing Test-Driven CodeEnables a granular scoring mechanism for tests of different difficulties to enhance policy optimization effectiveness.