Breeze-ASR-25 — MLX fp16

fp16 conversion of MediaTek-Research/Breeze-ASR-25 (a Whisper-large-v2 fine-tune for Taiwanese Mandarin, English, and Mandarin–English code-switching) into the mlx-community/whisper-* layout, for on-device use in SettleVoice via mlx-audio-swift.

Changes from upstream: weights cast to float16 and renamed to mlx_whisper keys (encoder.blocks.N.attn.query, decoder.positional_embedding, conv weights as [out, kernel, in]), encoder sinusoidal positions dropped, and a tokenizer.json generated from upstream vocab.json + merges.txt (upstream ships none). No weights were altered beyond the dtype cast. Model behaviour, limitations and license (Apache 2.0) are MediaTek Research's; see the upstream card.

Conversion script: scripts/convert_whisper_hf_to_mlx.py in the SettleVoice repo.

Downloads last month: -

Safetensors

Model size

2B params

Tensor type

F16

MLX

Hardware compatibility

Quantized

Model tree for PMTC/Breeze-ASR-25-mlx-fp16

Base model

openai/whisper-large-v2

Finetuned

MediaTek-Research/Breeze-ASR-25

Finetuned

(26)

this model