Xiaomi has introduced an open-source artificial intelligence (AI) model focused on reasoning, named MiMo. This innovative family of models aims to optimize reasoning abilities while maintaining a compact parameter size, marking the company’s first venture into open-source reasoning technology. MiMo will compete against notable Chinese models such as DeepSeek R1 and Alibaba’s Qwen QwQ-32B, in addition to global players like OpenAI’s o1 and Google’s Gemini 2.0 Flash Thinking. The MiMo series includes four distinct models, each designed for specific applications.
Xiaomi’s MiMo Reasoning AI Model to Compete With DeepSeek R1
The development of the MiMo series is a response to the significant size of existing reasoning AI models, which typically feature around 24 billion parameters or more. Such large sizes are intended to facilitate uniform advancements in both coding and mathematical capabilities across large language models—a challenging feat for smaller models.
In contrast to its competitors, MiMo consists of just seven billion parameters, with Xiaomi asserting that its performance aligns with that of OpenAI’s o1-mini and surpasses several reasoning models that have 32 billion parameters. The foundation model for MiMo was pre-trained on an impressive 25 trillion tokens.
Efficiency in MiMo’s development was achieved through the optimization of data preprocessing pipelines, the enhancement of text extraction toolkits, and the application of multidimensional data filtering techniques. Additionally, the pre-training of MiMo employed a three-stage data mixture strategy.
Internal tests conducted by Xiaomi researchers indicate that the MiMo-7B-Base scores 75.2 on the BIG-Bench Hard (BBH) benchmark, highlighting its reasoning capabilities. The zero-shot reinforcement learning (RL)-based MiMo-7B-RL-Zero reportedly excels in mathematical and coding tasks, obtaining a score of 55.4 on the AIME benchmark—4.7 points higher than that of o1-mini.
As an open-source initiative, the MiMo model is available for download through Xiaomi’s listings on GitHub and Hugging Face. The technical report provides comprehensive details on the model’s architecture, as well as its pre-training and post-training methodologies. It is important to note that MiMo is a text-based model and does not incorporate multimodal functionalities, with specifics about its training dataset remaining undisclosed.
