The MiMo-7B-RL model has demonstrated impressive performance in benchmark evaluations, achieving high Pass@1 rates in various mathematics and coding tasks.

The MiMo family consists of four distinct models, each designed for various applications, emphasizing efficient reasoning in a smaller parameter size.

The entire MiMo-7B model series is open-source and accessible on platforms like GitHub and Hugging Face, although specifics about its training dataset remain undisclosed.

Xiaomi has officially entered the AI industry with the launch of MiMo, its proprietary open-source AI model designed for reasoning capabilities.

Following the launch of MiMo, Xiaomi's shares rose over 5%, reflecting positive market sentiment and the broader trend of rising Chinese AI stocks.

This initiative aligns with China's broader strategy for self-sufficiency and innovation in AI technology, particularly in the face of increasing global competition.

In addition to its AI advancements, Xiaomi is also expanding into the electric vehicle market with the upcoming launch of its SU7 lineup.

Xiaomi's commitment to AI development is further evidenced by its significant investment in infrastructure, including the acquisition of approximately 10,000 GPUs for model training.

This move comes amid a rapidly developing AI landscape in China, where competition from companies like DeepSeek and Alibaba is driving innovation.

MiMo's effectiveness is largely attributed to its optimized pre-training process, which utilized a specialized dataset of 200 billion reasoning tokens and a three-stage data mixing strategy.

To enhance training efficiency, Xiaomi implemented a Seamless Rollout Engine, resulting in significant speed improvements in both training and validation processes.

Post-training reinforcement learning involved tackling over 130,000 verified math and programming problems, enhancing MiMo's reasoning skills through a custom Test Difficulty Driven Reward system.