Xiaomi has unveiled its open-source artificial intelligence (AI) model, known as MiMo, aimed specifically at enhancing reasoning capabilities while maintaining a smaller parameter size. This introduction marks the company’s first foray into open-source reasoning models, positioning MiMo against competitors like DeepSeek R1 and Alibaba’s Qwen QwQ-32B, in addition to global players such as OpenAI’s o1 and Google’s Gemini 2.0 Flash Thinking. The MiMo series consists of four distinct models, each tailored for specific applications.
Xiaomi’s MiMo Reasoning AI Model to Compete With DeepSeek R1
The motivation behind the MiMo series is to address the typical size constraints associated with reasoning AI models, which often exceed 24 billion parameters. Historically, larger models have sought to ensure simultaneous advancements in both coding and mathematical abilities, a challenge not easily met by smaller alternatives.
In contrast, MiMo boasts only seven billion parameters, with Xiaomi asserting that its performance is on par with OpenAI’s o1-mini and surpasses several models that have 32 billion parameters. The foundational AI model was reportedly pre-trained using an extensive dataset of 25 trillion tokens.
The researchers attribute this remarkable efficiency to several factors, including the optimization of data preprocessing processes, improvements to text extraction toolkits, and the implementation of multidimensional data filtering techniques. The pre-training phase utilized a three-stage data mixture strategy to further enhance model capabilities.
According to internal assessments, the MiMo-7B-Base achieves a score of 75.2 on the BIG-Bench Hard (BBH) benchmark, which measures reasoning skills. Additionally, the zero-shot reinforcement learning-based MiMo-7B-RL-Zero demonstrates exceptional performance in mathematics and coding, scoring 55.4 on the AIME benchmark, surpassing o1-mini by 4.7 points.
As an open-source offering, the MiMo model is available for download through Xiaomi’s official listings on GitHub and Hugging Face. The accompanying technical paper outlines the model’s architecture along with the methodologies employed in its pre-training and post-training phases. Notably, MiMo is a text-based model and lacks multimodal capabilities, and like many open-source projects, specific details regarding the dataset remain undisclosed.
