Maxkuo/Breeze-ASR-25-mlx-q8

Breeze-ASR-25 (MLX, 8-bit)

An 8-bit MLX conversion of MediaTek-Research/Breeze-ASR-25 for mlx-whisper on Apple Silicon. The weights are a straight quantization of the original model; nothing was fine-tuned.

MediaTek-Research/Breeze-ASR-25 的 MLX 8-bit 版，給 Apple Silicon 上的 mlx-whisper 用。只做格式轉換與量化，沒有再訓練。

Files

File	Contents
weights.safetensors	1,644,525,304 bytes, SHA-256 e4f8afaf5f5dc4c1a8d9e74e0101f5c66f4d6148db17b73d005fb753d071a8cf
config.json	Model dimensions and "quantization": {"group_size": 64, "bits": 8}
LICENSE	Apache License 2.0, copied unchanged from the base model

mlx-whisper 0.4.3 loads only weights.safetensors, so the converter's model.safetensors was renamed. The tokenizer is built into mlx-whisper; no tokenizer files are needed.

How it was made

• Base model: MediaTek-Research/Breeze-ASR-25 at revision cffe7ccb404d025296a00758d0a33468bec3a9d0
• Converter: whisper/convert.py from ml-explore/mlx-examples at commit e52c128d113f10546f0fa391f87edcc58d3880cb
• Environment: Python 3.12, mlx 0.32.4, mlx-whisper 0.4.3, torch 2.14.1, macOS 26.6.2 on Apple M5

python convert.py \
  --torch-name-or-path <local snapshot of MediaTek-Research/Breeze-ASR-25> \
  --mlx-path breeze-asr-25-q8 -q --q-bits 8 --q-group-size 64
mv breeze-asr-25-q8/model.safetensors breeze-asr-25-q8/weights.safetensors

The conversion is deterministic: running it twice (2026-10-25 and 2026-10-28) gave the same SHA-256, which also matches the 8-bit uploads at david20571015/Breeze-ASR-25-mlx-q8 and rueispaw/Breeze-ASR-25-mlx-q8.

Usage

import mlx_whisper

result = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="Maxkuo/Breeze-ASR-25-mlx-q8",
    language="zh",
    temperature=0.0,
)
print(result["text"])

Accuracy compared with fp16

Measured once on an Apple M5 MacBook Air (16 GB) with greedy decoding, language="zh", temperature=0.0:

Test set	fp16	8-bit
20 read sentences, Mandarin with English terms (387 tokens)	4.9%	4.9%
ASCEND test, every third utterance (439 utterances)	17.3%	17.3%
AliMeeting R8009_M8018, 27.6 min, CER	16.15%	16.17%

Error rates count one Chinese character or one English word as a token. These are small samples from one machine; check on your own audio before relying on them.

License

Apache License 2.0, the same as the base model. Changes from the base model: converted from the Hugging Face format to MLX and quantized the weights to 8 bits (group size 64).