mlx-community/MiMo-V2.5-ASR-MLX-8bit

Current variant: 8bit

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
MiMo-V2.5-ASR: Robust Speech Recognition Across
Languages, Dialects, and Complex Acoustic Scenarios
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

| 🤗 Official Model  | 🚀 Official Demo  | 📰 Official Blog  | 💻 Official Code  |

MLX Note

This repository is a community MLX conversion of the official XiaomiMiMo/MiMo-V2.5-ASR release for Apple silicon. The original model description below is preserved from the official release, and the MLX-specific material in this page is added as an incremental note for local MLX deployment.

MLX Usage

Current MLX usage is documented in the GitHub forks below:

• ailuntx/MiMo-V2.5-ASR-MLX
• ailuntx/MiMo-Audio-Tokenizer-MLX

Install the current MLX path:

pip install git+https://github.com/ailuntx/mlx-audio@feat/mimo-v25-asr

Download the MLX checkpoints:

hf download mlx-community/MiMo-Audio-Tokenizer --local-dir ./models/MiMo-Audio-Tokenizer
hf download mlx-community/MiMo-V2.5-ASR-MLX --local-dir ./models/MiMo-V2.5-ASR-MLX

Run transcription from the helper script in ailuntx/MiMo-V2.5-ASR-MLX:

git clone https://github.com/ailuntx/MiMo-V2.5-ASR-MLX.git
cd MiMo-V2.5-ASR-MLX
python run_mimo_asr_mlx.py \
    --model ./models/MiMo-V2.5-ASR-MLX \
    --audio path/to/audio.wav

Python:

from mlx_audio.stt import load

model = load("./models/MiMo-V2.5-ASR-MLX")
result = model.generate("path/to/audio.wav", language="en")
print(result.text)

Notes:

• mlx-community/MiMo-V2.5-ASR-MLX resolves mlx-community/MiMo-Audio-Tokenizer through mlx_manifest.json.
• The current install path depends on the MiMo support branch in ailuntx/mlx-audio.
• The usage section here will be simplified once MiMo lands in upstream mlx-audio and mlx-audio-swift.

Introduction

MiMo-V2.5-ASR is a state-of-the-art end-to-end automatic speech recognition (ASR) model developed by the Xiaomi MiMo team. It is built to deliver accurate and robust transcription across Mandarin Chinese and English, multiple Chinese dialects, code-switched speech, song lyrics, knowledge-intensive content, noisy acoustic environments, and multi-speaker conversations. MiMo-V2.5-ASR achieves state-of-the-art results on a wide range of public benchmarks.

Abstract

Automatic speech recognition systems are expected to faithfully transcribe speech signals that originate from diverse languages, dialects, accents, and domains, and that are captured under a wide variety of acoustic conditions. While conventional end-to-end models perform well on in-domain data, they still fall short of real-world requirements in challenging scenarios such as dialect mixing, code-switching, knowledge-intensive content, noisy environments, and multi-speaker conversations. Therefore, we present MiMo-V2.5-ASR, an end-to-end speech recognition model developed by the Xiaomi MiMo team. Through large-scale mid-training, high-quality supervised fine-tuning, and a novel reinforcement-learning algorithm, MiMo-V2.5-ASR achieves systematic improvements along the following dimensions:

• 🗣️ Chinese Dialects: Native support for Wu, Cantonese, Hokkien, Sichuanese, and more.
• 🔀 Code-Switch: Seamless Chinese-English code-switching transcription with no language tags required.
• 🎵 Song Recognition: High-precision lyrics transcription for Chinese and English songs, even with mixed accompaniment and vocals.
• 🔊 Noisy Environments: Robust recognition under heavy noise, far-field capture, and other adverse acoustic conditions.
• 👥 Multi-Speaker: Accurate transcription of overlapping, multi-party conversations such as meetings.
• 🇬🇧 Complex English Scenarios: Leading performance on the Open ASR Leaderboard for challenging English benchmarks such as AMI.
• 📚 Knowledge-Intensive Recognition: Precise recognition of classical poetry, technical terminology, personal names, place names, and other knowledge-dense material.
• 📝 Native Punctuation: Punctuation generated natively from prosody and semantics, delivering ready-to-use transcripts with no post-processing needed.

Results

MiMo-V2.5-ASR has been evaluated across a broad set of benchmarks spanning standard Mandarin and English, Chinese dialects, lyric recognition, and internal business scenarios. The chart below summarizes the average performance of MiMo-V2.5-ASR across these scenarios.

For per-benchmark numbers and specific qualitative cases, please refer to the official blog.

Model Download

Models	🤗 Hugging Face
MiMo-Audio-Tokenizer	XiaomiMiMo/MiMo-Audio-Tokenizer
MiMo-V2.5-ASR	XiaomiMiMo/MiMo-V2.5-ASR

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

MLX Releases

The following repositories are MLX conversions derived from the official release:

Variant	Precision	Size	Local smoke time	Smoke result
MiMo-V2.5-ASR-MLX	4bit	4.2 GB	0.88 s	Intention.
MiMo-V2.5-ASR-MLX-4bit	4bit	4.2 GB	0.88 s	Intention.
MiMo-V2.5-ASR-MLX-8bit	8bit	8.0 GB	10.80 s	Intention.
MiMo-V2.5-ASR-MLX-bf16	bf16	15 GB	-	dense reference export
MiMo-V2.5-ASR-MLX-fp32	fp32	30 GB	-	dense reference export

MLX conversion notes:

• Base model: XiaomiMiMo/MiMo-V2.5-ASR
• Tokenizer resolution: automatic via mlx-community/MiMo-Audio-Tokenizer
• Conversion date: 2026-05-12
• Local validation runtimes: mlx-audio and mlx-audio-swift
• Recommended default: MiMo-V2.5-ASR-MLX

Example downloads:

hf download mlx-community/MiMo-V2.5-ASR-MLX --local-dir ./models/MiMo-V2.5-ASR-MLX
hf download mlx-community/MiMo-V2.5-ASR-MLX-8bit --local-dir ./models/MiMo-V2.5-ASR-MLX-8bit

Validation

Local smoke validation was run with mlx-audio and mlx-audio-swift.

• intention.wav -> Intention.
• conversational_a.wav -> expected coffee / Kaldi paragraph

Getting Started

The following section is preserved from the official project and describes the original Python/CUDA workflow.

Spin up the MiMo-V2.5-ASR demo in minutes with the built-in Gradio app.

Prerequisites (Linux)

• Python 3.12
• CUDA >= 12.0

Installation

git clone https://github.com/XiaomiMiMo/MiMo-V2.5-ASR.git
cd MiMo-V2.5-ASR-MLX
pip install -r requirements.txt
pip install flash-attn==2.7.4.post1

If the compilation of flash-attn takes too long, you can download the precompiled wheel and install it manually:

• Download Precompiled Wheel

pip install /path/to/flash_attn-2.7.4.post1+cu12torch2.6cxx11abiFALSE-cp312-cp312-linux_x86_64.whl

Run the Demo

python run_mimo_asr.py

This launches a local Gradio interface for MiMo-V2.5-ASR. You can:

• Upload an audio file or record directly from your microphone.
• Optionally specify a language tag (Chinese / English / Auto) to bias the model for a specific language, or leave it to Auto for automatic language detection (recommended for code-switched speech).
• The demo calls the asr_sft() interface under the hood.

To load the model and tokenizer automatically at startup, pass their paths on the command line:

python run_mimo_asr.py \
    --model-path ./models/MiMo-V2.5-ASR \
    --tokenizer-path ./models/MiMo-Audio-Tokenizer

Otherwise, enter the local paths for MiMo-Audio-Tokenizer and MiMo-V2.5-ASR in the Model Configuration tab, then start transcribing.

Python API

The following API example is preserved from the official project.

Basic usage with the asr_sft interface:

from src.mimo_audio.mimo_audio import MimoAudio

model = MimoAudio(
    model_path="./models/MiMo-V2.5-ASR",
    tokenizer_path="./models/MiMo-Audio-Tokenizer",
)

# Automatic language detection (recommended for code-switching)
text = model.asr_sft("path/to/audio.wav")
print(text)

# With explicit language tag
text_zh = model.asr_sft("path/to/audio.wav", audio_tag="<chinese>")
text_en = model.asr_sft("path/to/audio.wav", audio_tag="<english>")

Citation

@misc{coreteam2026mimov25asr,
      title={MiMo-V2.5-ASR: Robust Speech Recognition Across Languages, Dialects, and Complex Acoustic Scenarios},
      author={LLM-Core-Team Xiaomi},
      year={2026},
      url={https://github.com/XiaomiMiMo/MiMo-V2.5-ASR},
}

Contact

Please contact mimo@xiaomi.com or open an issue in the official project if you have questions about the original model.