MiMo-V2.5-ASR MLX INT4

This repository contains the MLX int4 weight-only quantized ASR checkpoint generated from the local bf16 MiMo-V2.5-ASR package.

Files

  • model.safetensors
  • model.safetensors.index.json
  • config.json with quantization_config
  • Qwen tokenizer/config sidecar files

Load

from mimo_mlx import load_asr

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

The audio tokenizer is required separately. Use the official XiaomiMiMo/MiMo-Audio-Tokenizer files or the local directory expected by the project.

Fallback

Keep the bf16 package available for quality fallback:

from mimo_mlx import load_asr

asr = load_asr("bf16")

Validation

On the local VoiceInk-History seeded 10-sample spread check, this int4 package kept the fixed sample exact and did not show degradation versus bf16 on that small check.

Source

Base model: https://huggingface.co/XiaomiMiMo/MiMo-V2.5-ASR

Downloads last month
296
Safetensors
Model size
1B params
Tensor type
BF16
·
U32
·
MLX
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for carloshuang1224/MiMo-V2.5-ASR-MLX-INT4

Quantized
(5)
this model