Xiaomi released an artificial intelligence (AI) model based on open source reasoning on Tuesday. Dubbed memo, the family’s family’s relatively smaller innovation in the small parameter size improves the reasoning capacity of reasoning. It is also a model of the first open source reasoning through Tech Dev, and it competes with global reasoning, including Chinese models such as Dipic R1 and Alibaba’s Kevin QWQ -32B, and Openi’s O -1 and Google’s Gemini 2.0 flash thinking. The MIMO family includes four different models, each of which are unique issues of use.
Xiaomi’s MIMO Request to compete with AI Model DepsEek R1
With the MIMO series of AI models, Xiaomi researchers aim to solve the size problem in the AI model argument. There are about 24 billion or more parameters (at least which can be measured). Large size is kept to achieve the same and simultaneous improvement in both coding and math capabilities of large language models, which is considered difficult to achieve with small models.
In comparison, seven billion parameters have been presented in the MIMO, and Xiaomi claims that its performance is similar to the Openi O-1-Mani and improves multiple reasoning models with 32 billion parameters. Researchers claimed that the base AI model was already trained on 25 trillion tokens.
Researchers claimed that such performance has been achieved by pre -processing pipelines, increasing text extraction tool kits, and applying multi -dimensional data filtering. In addition, the memo -pre -training includes a three -step data mixture strategy.
Based on internal testing, Xumi researchers claim that the MIMO 7B Base has scored 75.2 on the Big Bench Hard (BBH) benchmark for reasoning capabilities. Zero Shot Kimk Learning (RL) based on MIMO-7B-RL Z Zero Mathematics and coding-related tasks are claimed to perform well, and the AIME scores 55.4 on the benchmark, which performs better than 4.7 points to O1-MINI.
Since MIMO is an open source AI model, it can be downloaded from the Xiaomi list Got hub And The hugs face. Technical Paper Details of the model’s architecture as well as training before training and after training. It is a text -based model and does not have multi -modal capabilities. Like most open source release, details about the model dataset are not known.
