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-V2-Flash

MiMo-V2-Flash

High-speed MoE language model for reasoning and agents

Xiaomi released MiMo-V2-Flash, a mixture of experts model with 309 billion total parameters. It uses hybrid attention and multi-token prediction to speed up inference while supporting long contexts.

Share on WhatsApp

  • 309 billion total parameters with 15 billion active
  • Supports context lengths up to 256,000 tokens
  • Uses hybrid attention to reduce memory usage
  • Multi-token prediction triples generation speed
  • Trained on 27 trillion tokens of data

What it is

MiMo-V2-Flash is a large language model built by Xiaomi. It contains 309 billion parameters in total, but only activates 15 billion for each token it processes. This design aims to balance high performance with lower computational cost during use. The model supports a context window of up to 256,000 tokens, allowing it to handle very long documents or codebases. It is designed specifically for complex reasoning tasks and autonomous agent workflows.

How it came about

The model was developed by Xiaomi as part of its MiMo series. It represents an advancement in efficient large-scale modelling by combining mixture of experts architecture with novel training techniques. The developers focused on creating a system that could perform well on standard benchmarks while remaining fast enough for practical deployment in agentic settings.

Inside the model

The architecture uses a hybrid attention mechanism that mixes sliding window attention with global attention. Most layers use a small window of 128 tokens to save memory, while occasional layers look at the entire sequence. A learnable bias helps maintain accuracy despite the small windows. The model also includes a multi-token prediction module that generates several tokens at once, tripling output speed. Training involved 27 trillion tokens using mixed precision methods.

What it can do

The base model performs strongly on general knowledge tests like MMLU and coding benchmarks such as HumanEval. After post-training, it shows improved results in mathematics and software engineering tasks. Evaluations indicate it competes with other leading models on agent-based benchmarks like SWE-Bench. It handles multiple languages including English and Chinese effectively across various difficulty levels.

How to use it

Xiaomi has made the model weights available to the public via Hugging Face. Both the base version and the post-trained version can be downloaded. The company also released the weights for the multi-token prediction module to support further research by the community. Users can run the model locally if they have sufficient hardware resources.

Questions people ask about MiMo-V2-Flash

How many parameters does MiMo-V2-Flash have?

The model has 309 billion total parameters. However, it is a mixture of experts model, so it only uses 15 billion active parameters for each token it processes.

What makes this model faster than others?

It uses a multi-token prediction module that generates multiple tokens simultaneously. This technique triples the output speed during inference compared to standard single-token generation methods.

Can it handle long documents?

Yes, the model supports a context window of up to 256,000 tokens. It uses hybrid attention with sliding windows to manage this length efficiently without excessive memory costs.

Is the model open source?

The model weights are open and available for download on Hugging Face. Xiaomi also released the weights for the multi-token prediction module to help researchers study the architecture.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on MiMo-V2-Flash yet, so the facts on this page come from the sources listed above: huggingface.co/XiaomiMiMo/MiMo-V2-Flash. 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.