News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
MiMo-7B

MiMo-7B

Open reasoning model from Xiaomi for maths and code

Xiaomi released MiMo-7B, a series of open-weight language models designed for reasoning tasks. Trained on 25 trillion tokens, the model uses reinforcement learning to match larger systems in mathematics and coding benchmarks.

Share on WhatsApp

  • Trained on 25 trillion tokens for reasoning
  • Uses Multiple-Token Prediction for faster inference
  • Scores 95.8 on MATH-500 benchmark
  • Open weights available on Hugging Face
  • Reinforcement learning uses rule-based verifiers

What it is

MiMo-7B is a family of language models built by Xiaomi for complex reasoning tasks. The series includes base, supervised fine-tuning, and reinforcement learning variants. It focuses on improving performance in mathematics and computer programming without relying on massive parameter counts. The developers claim the model matches the capabilities of larger proprietary systems through specialised training strategies.

How it came about

The project aims to unlock reasoning potential from pre-training rather than just post-training adjustments. Xiaomi trained the base model on a vast dataset and then applied reinforcement learning to specific checkpoints. In May 2025, the team updated the supervised fine-tuning data and expanded the training window. This update allowed the model to surpass other open-source competitors on specific mathematics benchmarks.

Inside the model

The base model underwent pre-training on approximately 25 trillion tokens using a three-stage data mixture. Developers optimised text extraction and filtering to increase reasoning patterns. They employed Multiple-Token Prediction to speed up inference and improve accuracy. For post-training, the team used rule-based verifiers on 130,000 problems to avoid reward hacking. A custom rollout engine accelerated the reinforcement learning process significantly.

What it can do

The model demonstrates strong results in mathematics and coding. On the MATH-500 benchmark, it achieved a score of 95.8. It scored 68.2 on AIME 2024 and 57.8 on LiveCodeBench v5. These figures place it close to or above several larger models in these specific domains. The developers state it matches the performance of OpenAI o1-mini in reasoning tasks.

How to use it

Xiaomi provides open access to the model weights via Hugging Face and ModelScope. The release includes checkpoints for the base model, the supervised fine-tuned version, and two reinforcement learning variants. Users can run the model using SGLang or a specific fork of vLLM that supports its architecture features.

Questions people ask about MiMo-7B

Who developed MiMo-7B?

Xiaomi developed the MiMo-7B series. The company released the models as open weights to support the broader community in building powerful reasoning language models.

What makes MiMo-7B different from other small models?

It is designed specifically for reasoning from the pre-training stage. It uses Multiple-Token Prediction and a specialised reinforcement learning recipe with rule-based rewards to enhance maths and code skills.

How can I run MiMo-7B?

You can download the weights from Hugging Face or ModelScope. The model supports inference through SGLang and a custom fork of vLLM that handles its Multiple-Token Prediction layers.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on MiMo-7B yet, so the facts on this page come from the sources listed above: huggingface.co/XiaomiMiMo/MiMo-7B-RL. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.