mimo-mlx 0.1.0

MLX-native MiMo-V2.5-ASR inference package for Apple Silicon

mimo-mlx

MLX-native MiMo-V2.5-ASR inference package for Apple Silicon.

This package contains the Python inference code only. Model weights are hosted on Hugging Face as separate repositories:

• INT4: https://huggingface.co/carloshuang1224/MiMo-V2.5-ASR-MLX-INT4
• BF16: https://huggingface.co/carloshuang1224/MiMo-V2.5-ASR-MLX-BF16
• F32: https://huggingface.co/carloshuang1224/MiMo-V2.5-ASR-MLX-F32

The MiMo audio tokenizer is required and is downloaded from https://huggingface.co/XiaomiMiMo/MiMo-Audio-Tokenizer when using download=True.

Install

pip install mimo-mlx

Quick Start

from mimo_mlx import load_asr

asr = load_asr("int4", download=True)
text = asr.transcribe("audio.wav", language="zh")
print(text)

Use bf16 for the validated fallback path:

from mimo_mlx import load_asr

asr = load_asr("bf16", download=True)

If the model directories are already present locally:

from mimo_mlx import load_asr

asr = load_asr("int4")

Expected local directory layout:

models/
  MiMo-Audio-Tokenizer/
  MiMo-V2.5-ASR-int4/
  MiMo-V2.5-ASR-bf16/

Notes

• The INT4 package uses MLX quantized safetensors and requires the quantized load path in this package.
• The default MiMoASR() constructor uses local bf16 paths for quality fallback.
• Model weights are not bundled in the PyPI package.