BRlin/Breeze-ASR-25-mlx-bf16

Breeze-ASR-25 · MLX — superseded

Use BRlin/Breeze-ASR-25-mlx-fp16 instead. A 4-bit build is also available at BRlin/Breeze-ASR-25-mlx-4bit.

This repository remains available so that existing references keep working. It is not being updated.

Why it was replaced

The name is wrong. The weights here are float16, not bfloat16. They were converted with mlx_audio.convert --dtype bfloat16, which prints [INFO] Converting to bfloat16 but does not apply the requested dtype on the non-quantizing path — the cast is written to a variable that is discarded before the file is saved. The mistake was ours for trusting the flag rather than inspecting the output.

float16 is in fact the correct target for this model: mlx-whisper accepts only float16 or float32, and rejects bfloat16 outright. So the weights themselves were never the problem.

The layout served only one runtime. config.json here carries the HuggingFace transformers schema, which mlx-whisper cannot parse — it expects its own ten-field ModelDimensions schema and raises TypeError on anything else. This repository therefore loads in mlx-audio but not in mlx-whisper.

What the replacement changes

The weights in BRlin/Breeze-ASR-25-mlx-fp16 are byte-identical to the ones here. If you already have this repository, you have the correct model; you do not need to re-download for correctness.

What changed is the packaging and the documentation:

• config.json rewritten into the MLX schema, so both mlx-whisper and mlx-audio load it
• the name states the actual dtype
• evaluation against the figures the upstream authors publish, on public datasets, with the scoring method stated