Xiaomi MiMo
AI Model for Precise Reasoning and Efficient Design
MiMo is an open-source artificial intelligence model developed by Xiaomi that specializes in mathematical reasoning and code generation. It integrates advanced architecture with data optimization to efficiently solve complex problems with high performance and a compact size.
161
Social networks of Xiaomi MiMo
AI categories of Xiaomi MiMo
Large Language Model (LLM) Building
¡Producto agotado!
What is Xiaomi MiMo ?
MiMo-7B is a language model with 7 billion parameters developed by Xiaomi, focused on mathematical reasoning and programming tasks. This model combines an advanced architecture with specific techniques to maximize its capability to solve complex problems despite its relatively compact size.
The creation process of MiMo-7B consists of two main stages. In the first stage, it was trained with a carefully selected data collection including academic content, technical texts, and mathematical problems. 70% of this data focused on mathematics and programming. The model processed 25 trillion tokens during this phase and used multiple prediction techniques to improve its efficiency.
The second stage refined the model through two methods: supervised learning with 500,000 examples and reinforcement learning with 130,000 verifiable practical problems. For this latter process, a system was developed that automatically evaluates the correctness of solutions proposed by the model and trains it to progressively improve. The training infrastructure was optimized to accelerate this process more than twofold.
In standardized evaluations, MiMo-7B achieved notable scores in reasoning tests (BBH: 75.2), mathematical competitions (AIME: 55.4), and programming (LiveCodeBench v5: 57.8). These results demonstrate its competitiveness against models of similar size and even some larger ones.
The model is available in four versions (Base, SFT, RL-Zero, and RL) representing different stages of its development.
The creation process of MiMo-7B consists of two main stages. In the first stage, it was trained with a carefully selected data collection including academic content, technical texts, and mathematical problems. 70% of this data focused on mathematics and programming. The model processed 25 trillion tokens during this phase and used multiple prediction techniques to improve its efficiency.
The second stage refined the model through two methods: supervised learning with 500,000 examples and reinforcement learning with 130,000 verifiable practical problems. For this latter process, a system was developed that automatically evaluates the correctness of solutions proposed by the model and trains it to progressively improve. The training infrastructure was optimized to accelerate this process more than twofold.
In standardized evaluations, MiMo-7B achieved notable scores in reasoning tests (BBH: 75.2), mathematical competitions (AIME: 55.4), and programming (LiveCodeBench v5: 57.8). These results demonstrate its competitiveness against models of similar size and even some larger ones.
The model is available in four versions (Base, SFT, RL-Zero, and RL) representing different stages of its development.
Related news
-
MiMo-7B: Xiaomi’s new AI model excels at mathematical reasoning
30/04/2025Xiaomi introduces MiMo-7B, an open-source language model that, with only 7 billion parameters, outperforms larger models in complex mathematical reasoning and programming tasks.Ampliar