Downloads · 30 days
43
10% of all-time downloads
cohere-transcribe-arabic-07-2026-MLX-Mixed-2bit3bit4bit
MarkChen1214/cohere-transcribe-arabic-07-2026-MLX-Mixed-2bit3bit4bit
cohere-transcribe-arabic-07-2026-MLX-Mixed-2bit3bit4bit is a automatic speech recognition model from MarkChen1214. Use it when you need speech turned into text. It is set up for mlx. The card lists the license as mit.
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.
All-time downloads
419
Public
Parameters
2.1B
842 MB on disk
Likes
1
Public
Repo makeup
Click a slice to open those files.
.safetensors842 MB · 100%
Parameter types
How the weights are stored.
U322.1B · 100%
Base models
Model card
From the Hugging Face model README
Cohere Transcribe Arabic 07-2026 — MLX Mixed 2-bit/3-bit/4-bit
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.
Key Metrics
| Metric | Value |
|---|---|
| Size | 806 MB (vs 4.13 GB FP16 — 5.1x smaller) |
| Effective bits/param | 3.456 |
| WER FLEURS ar_eg (full 428) | 5.09% (FP16: 4.74%) |
| WER LibriSpeech clean/other (full) | 1.14% / 2.55% (composite 1.84%) |
| RTFx (M4 Air, MLX) | ~19x on Arabic — 2x the FP16 speed (9.8x) |
Arabic WER is scored with standard Arabic normalization (diacritics stripped, hamza-alif/ya/ta-marbuta folded); compare deltas, not absolute numbers across papers.
Quantization layout
| Component | Quantization |
|---|---|
| Encoder (most groups) | 2-bit affine, group size 64 |
| Encoder blocks 36–47 (attn/ffn/pointwise, per bitmap) | 3–4-bit |
| Decoder + head | 4-bit |
| Norms / embeddings / mel frontend | FP16 |
1x1 Conv1d layers are converted to Linear equivalents to enable quantization.
Architecture
- Base model: Cohere Transcribe Arabic 07-2026 (~2B params, bilingual ar+en)
- Encoder: FastConformer (48 layers, d=1280)
- Decoder: Transformer (8 layers, d=1024)
- Tokenizer: SentencePiece (16,384 tokens)
Usage
Requires mlx-audio installed from git main:
pip install "mlx-audio[stt] @ git+https://github.com/Blaizzy/mlx-audio.git"
# apply the loader patch shipped in this repo (fixes two upstream
# quantized-reload bugs; see mlx_audio_cohere_quant_patch.py)
from mlx_audio_cohere_quant_patch import apply_patch
apply_patch()
from mlx_audio.stt import load
model = load("MarkChen1214/cohere-transcribe-arabic-07-2026-MLX-Mixed-2bit3bit4bit")
# Arabic (incl. dialects and ar-en code-switching)
result = model.generate(audio="audio.wav", language="ar", punctuation=True)
print(result.text)
# English works too
result = model.generate(audio="audio_en.wav", language="en", punctuation=True)
Eval Results (full datasets, seed=42)
| Dataset | Samples | WER | RTFx |
|---|---|---|---|
| FLEURS ar_eg test | 428 | 5.09% | 19.0x |
| LibriSpeech test-clean | 2,620 | 1.14% | 17.9x |
| LibriSpeech test-other | 2,939 | 2.55% | 18.9x |
English scored with the HF Open ASR Leaderboard methodology (Whisper EnglishTextNormalizer + jiwer).
License
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.
- Task
- Automatic Speech Recognition
- Library
- mlx
- Type
- cohere_asr
- License
- mit
- Languages
- ar, en
- Created
- Jul 15, 2026
- Updated
- Jul 18, 2026