Xiaomi on Tuesday released an open-source Reasoning-centric Artificial Intelligence (AI) model. Dubbed Mimo, the family of the Reasoning model innovates the optimization of logic capacity in relatively small parameters size. It is also the first open -source region model by the tech giant, and it competes with the Chinese models such as Dipsek R1 and Alibaba’s Quven QWW -32B, and OpenaiI’s O1 and Google’s Gemini 2.0 Flash Thinking with global logic models. The MIMO family consists of four separate models, each of which has unique use cases.
Xiaomi’s MIMO argued to compete with AI model Deepsek R1
With the MIMO series of the AI model, Xiaomi researchers aims to solve the problem of size in arguing the AI model. The logic model (at least that can be measured) contains about 24 billion or more parameters. The larger size is placed to get equal and simultaneous improvement in both coding and mathematical abilities of large language models, it is considered difficult to get with some small models.
In comparison, MIMO has seven billion parameters, and Xiaomi claims that its performance corresponds to OPENAI’s O1-Mini and improves several logic models with 32 billion parameters. Researchers claimed that the Base AI model was pre-educated on 25 trillion tokens.
Researchers claimed that such efficiency was achieved by adapting data preprosaaceing pipelines, enhancing text extraction toolkit and implementing multidimensional data filtering. In addition, the pre-training of MIMO included a three-step data mix strategy.
Depending on the internal test, Xiaomi researchers claim that MIMO-7B-Base has scored 75.2 score 75.2 for logic capabilities on large bench hard (BBH) benchmarks. Zero-SOT reinforcement Learning (RL)-Appeared MIMO-7B-RL-Zzero claims to excel in mathematics and coding-related tasks, and 55.4 scores on the AIME benchmark, O1-Min have been improved by 4.7 points.
As Mimo is an open-source AI model, it can be downloaded from the listing of Xiaomi to Github and Hugging Face. The technical paper gives details of the architecture of the model as well as the processes after pre-training and training. It is a lesson-based model and does not have multimodal capabilities. Similar to most open-sources release, details about the model’s dataset are not known.
