metadata
license: mit
language:
- en
- pl
- de
- fr
- es
- it
- pt
- nl
- ru
- zh
- ja
- ko
- ar
- hi
- multilingual
tags:
- whisper
- speech-to-text
- mlx
- quantized
- q4
- 4bit
- apple-silicon
library_name: transformers
pipeline_tag: automatic-speech-recognition
base_model: openai/whisper-large-v3
Whisper Large V3 - MLX Q4 Quantized
4-bit quantized version of OpenAI's Whisper Large V3, optimized for Apple Silicon with MLX. Smallest large-v3 variant.
Model Details
| Property | Value |
|---|---|
| Original Model | openai/whisper-large-v3 |
| Parameters | ~1.55B |
| Quantization | INT4 (Q4) |
| Size | ~900MB |
| Decoder Layers | 32 |
Half the size of Q8 with minimal accuracy loss. Best choice for memory-constrained devices.
Other Whisper Models
| Model | Size | Quality | Link |
|---|---|---|---|
| small | ~300MB | Good | LibraxisAI/whisper-small-mlx-q8 |
| medium | ~800MB | Better | LibraxisAI/whisper-medium-mlx-q8 |
| large-v3 q8 | ~1.6GB | Best | LibraxisAI/whisper-large-v3-mlx-q8 |
| large-v3 q4 (this) | ~900MB | Best (compressed) | - |
| large-v3-turbo | ~900MB | Great (fast) | LibraxisAI/whisper-large-v3-turbo-mlx-q8 |
Usage
import mlx_whisper
result = mlx_whisper.transcribe(
"audio.wav",
path_or_hf_repo="LibraxisAI/whisper-large-v3-q4"
)
print(result["text"])
Supported Languages
Full multilingual support: English, Polish, German, French, Spanish, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean, Arabic, Hindi, and 90+ additional languages.
Hardware Requirements
- Apple Silicon Mac (M1/M2/M3/M4)
- Minimum 8GB RAM
License
MIT - inherited from OpenAI Whisper.
Converted by LibraxisAI using mlx-whisper