flo5k5/Voxtral-Mini-4B-Realtime-2602-MLX-6bit
Automatic Speech Recognition • 4B • Updated • 8 • 1
Voxtral on Apple Silicon with mlx-audio: tested conversions and the original models.
Note 6-bit, mlx-audio format. Keeps up with live speech on an M5 Pro (text ~0.1 s behind the voice); same French WER as fp16.
Note Original streaming model by Mistral AI (Apache 2.0).
Note 8-bit, mlx-audio format. Keeps up with live speech on an M5 Pro and matches fp16 French WER; 1 GB heavier than the 6-bit.
Note 4-bit, mlx-audio format (lighter, not measured here).
Note fp16 source of the 6-bit conversion; slower than real time on an M5 Pro (0.47x).