BRlin/Breeze-ASR-25-ggml

Breeze-ASR-25 — ggml (whisper.cpp)

ggml conversions of MediaTek-Research/Breeze-ASR-25 for whisper.cpp. Breeze-ASR-25 is fine-tuned from Whisper large-v2, so whisper.cpp's large-v2 Core ML encoder works with these files if you use Core ML.

File	Type	Size
ggml-breeze-f16.bin	f16	3.09 GB
ggml-breeze-q8_0.bin	q8_0	1.66 GB
ggml-breeze-q5_0.bin	q5_0	1.08 GB
ggml-breeze-q4_1.bin	q4_1	0.98 GB
ggml-breeze-q4_0.bin	q4_0	0.89 GB

How they were made

whisper.cpp at commit 2ca53bb (2026-07-31):

python3 whisper.cpp/models/convert-h5-to-ggml.py <Breeze-ASR-25> <openai/whisper repo> <out>
whisper-quantize ggml-breeze-f16.bin ggml-breeze-<type>.bin <type>

Usage

whisper-cli -m ggml-breeze-q5_0.bin -l zh -f audio.wav

MLX versions of the same model: BRlin/Breeze-ASR-25-mlx-fp16, BRlin/Breeze-ASR-25-mlx-4bit. The license follows the base model (Apache-2.0).