Skip to main content
Rase AI Intelligence Newsletter

Models & assistants Model or AI service

Xiaomi’s MiMo-7B tests reasoning in a smaller language model

Xiaomi released downloadable MiMo-7B models for research on reasoning in mathematics and programming. The project distinguishes base, instruction-tuned and reinforcement-trained checkpoints. [1][2][3]

Model access
Original MiMo-7B checkpoints are publicly downloadable. [2][3]
License
The original MiMo-7B-RL model card lists MIT. The project repository carries Apache 2.0. [2][3]

What the model does

Xiaomi first submitted the MiMo paper in May 2025. It describes the 7B series and separates results for the base model from checkpoints trained further for reasoning. [1][2]

Why it matters

Several checkpoints of the same size allow researchers to compare how later training affects answers. That makes it easier to interpret differences than comparing models that also differ greatly in size. [1][2]

Who can use it

Researchers can download a checkpoint to study mathematical reasoning or how the model handles code. The model card identifies the original reinforcement-trained version. [2][3]

Key dates

  1. MiMo technical report first submitted [1]

Ways to use it

Give the base and reinforcement-trained checkpoints the same small set of questions with known answers. Record which build produced each answer so later variants stay out of the comparison. [1][3]

Model limitations

The reported strengths are tied to the checkpoints and test sets the researchers evaluated. The original RL model card labels its weights MIT, while the repository code uses Apache 2.0. [2][3]

Sources and verification

Catalog evidence checked How we verify records

  1. Primary source Xiaomi’s first MiMo technical report and submission history
  2. Primary source Xiaomi’s official MiMo model lineup and code repository
  3. Primary source XiaomiMiMo/MiMo-7B-RL
  4. Other source MiMo: Unlocking the Reasoning Potential of Language Model -- From Pretraining to Posttraining

Related profiles

Models & assistants

GPT-6.2 Astra and Sol

OpenAI

Astra targets demanding projects. Sol targets everyday tasks. Both support text and image input in the API.

High confidenceEvidence checked

Models & assistants

MAI Code 2 and Code 2 Flash

Microsoft AI

Code 2 targets complex software work. Flash targets frequent edits, explanations and routine fixes.

High confidenceEvidence checked

Models & assistants

MAI Realtime 1

Microsoft AI

Users can try a MAI Labs demo. Developers can register for a limited Foundry preview.

High confidenceEvidence checked