Xiaomi releases MiMo-7B

Release

In late April 2025, Xiaomi released the MiMo-7B series ^[1]. The release covered two versions: a pre-trained base model and a reinforcement learning (RL)-tuned chat model ^[1]. That pairing is common in model releases because it serves two audiences. The base model is the starting point for further training and experimentation; the chat model is tuned to follow instructions and solve problems in conversation. For people who follow reasoning models, the interesting part is not only the name but the training recipe and the reported results.

Training recipe

Xiaomi trained the base model on 25 trillion tokens ^[1]. Token count is a rough measure of how much text and code the model sees during pre-training; a larger number means more training data, though it says nothing by itself about data quality or model quality. For post-training, Xiaomi used 130K verifiable mathematics and programming problems ^[1]. Verifiable here means the answer can be checked automatically—for example, a math answer or a piece of code with tests—so the training process can reward correct outputs and penalize wrong ones. This is a practical way to use reinforcement learning on reasoning tasks, because the model can practice and receive a signal without a human grading every attempt. Xiaomi reported that MiMo-7B-RL beat OpenAI o1-mini ^[1]. That is a company-reported comparison, so it is best read as a claim about the model's intended position rather than as a final, independent ranking.

Release details and later revision

The Hugging Face model card was created on 29 April 2025 ^[2]. Model cards matter because they give developers a fixed place to find the model's intended use, training details, and evaluation notes. A 30 May 2025 revision pushed the reported AIME 2024 score to 80.1, above DeepSeek R1's 79.8 ^[3]. Benchmark numbers like these are useful for quick comparison, but they are also narrow: a high score on one test does not automatically mean the model is better at every task. The revision shows that the project's published results continued to change after the first release, which is common for models that are still being tuned and evaluated.

Why it matters

MiMo-7B is notable less as a single score than as a sign of how model releases are presented. A company can now ship a base model and a tuned chat model together, describe a post-training pipeline built on verifiable problems, and place its results next to established names such as OpenAI o1-mini and DeepSeek R1 ^[1][3]. For students and developers, the release is a useful case study in how reasoning models are trained and evaluated: data scale, reinforcement learning, model cards, and benchmark comparisons all appear in one package. The right response is to look at the recipe and the evidence, not just the headline comparison.