A
dynamically quantized MLX build of
CohereLabs/cohere-transcribe-arabic-07-2026
(bilingual Arabic + English, dialects, code-switching). To our knowledge the
first sub-1GB quantized release of this model.
Bit-widths are assigned per module from
two independent quantization-
sensitivity scans (one on this model, one on its English base model). The
published allocation keeps the safer, higher bit-width for each module. Most
of the encoder runs at
2-bit; the empirically fragile groups (decoder, last
encoder blocks) stay at 3/4-bit. The rule map ships in this repo as
quant_bitmap.json.
Arabic WER is scored with standard Arabic normalization (diacritics stripped,
hamza-alif/ya/ta-marbuta folded); compare deltas, not absolute numbers across
papers.
1x1 Conv1d layers are converted to Linear equivalents to enable quantization.
1# apply the loader patch shipped in this repo (fixes two upstream
2# quantized-reload bugs; see mlx_audio_cohere_quant_patch.py)
3from mlx_audio_cohere_quant_patch import apply_patch
4apply_patch()
5
6from mlx_audio.stt import load
7
8model = load("MarkChen1214/cohere-transcribe-arabic-07-2026-MLX-Mixed-2bit3bit4bit")
9
10# Arabic (incl. dialects and ar-en code-switching)
11result = model.generate(audio="audio.wav", language="ar", punctuation=True)
12print(result.text)
13
14# English works too
15result = model.generate(audio="audio_en.wav", language="en", punctuation=True)
English scored with the HF Open ASR Leaderboard methodology (Whisper
EnglishTextNormalizer + jiwer).
MIT — use it however you like, attribution appreciated.
The base model (
CohereLabs/cohere-transcribe-arabic-07-2026)
is Apache 2.0; its notices apply to the underlying weights.
