MiMo-V2.5-ASR

MiMo-V2.5-ASR

Transcribes speech in Mandarin, English and Cantonese, including Chinese dialects, code-switching, song lyrics, noisy audio and multi-speaker meetings, with punctuation included. For teams that need speech to text.

Made by XiaomiMiMo on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MiMo-V2.5-ASR is a 7.6B-parameter speech recognition model from the Xiaomi MiMo team. It turns audio into text for Mandarin Chinese, English and Cantonese. The card says it also handles several Chinese dialects, mixed Chinese and English speech, song lyrics and hard audio such as noise or overlapping voices.

Who it suits: It suits teams that transcribe Chinese and English audio, especially messy everyday recordings, and researchers who compare speech models.

Industry | Who uses it | How they use it | Example
Media and entertainment | Subtitle and lyrics teams | Transcribe songs and spoken video in Chinese or English | Getting lyrics from a song with music under the voice
Business | Meeting note takers | Transcribe multi-person conversations | Turning a recorded team meeting into a transcript
Customer service | Call analysis teams | Transcribe noisy or far-field recordings | Reading calls recorded in a busy room
Research | Speech researchers | Test dialect and code-switch recognition | Checking how it handles Sichuanese or Hokkien speech

You can skip it if you only need a small model on a laptop, since this one has 7.6B parameters and the card's demo targets a Linux GPU setup.

Where to use it:

• Python code that uses the transformers library
• Linux with Python 3.12 and CUDA 12.0 or newer, which the card lists for its demo
• A built-in Gradio demo app that comes with the project
• Windows, macOS or Linux, with Python for the file download

Use case | Who | What they get back
Transcribe dialect speech | Regional content teams | Text for Wu, Cantonese, Hokkien, Sichuanese and more
Transcribe code-switched speech | Bilingual speakers | Chinese and English text without language tags
Transcribe song lyrics | Music teams | Lyrics from songs with mixed vocals and music
Transcribe meetings | Business users | Text from overlapping speakers, with punctuation

How to set it up:

• Python 3.12 and CUDA 12.0 or newer on Linux, per the card
• The huggingface-hub package
• Disk space, about 30.5 GB for the full-precision weights
• The audio tokenizer model, which is downloaded alongside the speech model

1. Install the Hugging Face hub package with pip.
2. Download the audio tokenizer and the speech model into a local models folder.
3. Follow the card's installation steps and start its Gradio demo.

pip install huggingface-hub
hf download XiaomiMiMo/MiMo-Audio-Tokenizer --local-dir ./models/MiMo-Audio-Tokenizer
hf download XiaomiMiMo/MiMo-V2.5-ASR --local-dir ./models/MiMo-V2.5-ASR

How to use it: You start the demo app that comes with the project, load an audio file and read the transcript. The part of the card shown here does not include the install and run commands for the app.

Example: A person records a team meeting where two people talk over each other and switch between Chinese and English. They run the file through the model. What comes back is one transcript with punctuation already added, so it needs no extra cleanup step.

For newcomers and experts: If you are new to this, begin with one short, clear recording in Mandarin or English and check the transcript against what you hear. Move to dialect, song or noisy audio after that.

If you know the field, the card describes an end-to-end model built with large-scale mid-training, supervised fine-tuning and a new reinforcement learning method. It says punctuation comes from prosody and meaning, and that it leads on a public English leaderboard for hard sets like AMI. Per-benchmark numbers are on the authors' blog, not in the part shown here.

Download the tokenizer and the model
Install the project requirements
Start the Gradio demo
Load an audio file
Read the punctuated transcript

Good to know:

• Limits: the card summarizes results in a chart and points to a blog for numbers, and the install commands are cut from the part shown here.
• Claims: the quality statements come from the authors.
• Hardware: the card lists Linux with Python 3.12 and CUDA 12.0 or newer for its demo.
• License: MIT.
• Cost to run: none named.
• Data it reaches: it copies two model downloads to your computer and reads the audio files you give it.

On Hugging Face, 13 November 2026

• Task: automatic speech recognition
• Library: transformers
• Parameters: 7.6B
• Weights: about 30.5 GB in F32; about 3.8 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 4,626
• Likes: 117
• License: mit
• Languages: zh, en, yue
• Quantized versions: 6
• Fine-tunes: 6
• Tags: transformers, safetensors, qwen2, text-generation, automatic-speech-recognition

Links

• Model: https://huggingface.co/XiaomiMiMo/MiMo-V2.5-ASR
• Files: https://huggingface.co/XiaomiMiMo/MiMo-V2.5-ASR/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download XiaomiMiMo/MiMo-V2.5-ASR

Then follow the usage in its model card. Nothing in this kit runs by itself.