whisper-large-v3-turbo — MLX int8

Files

• weights.safetensors — int8 quantized weights (~864 MB)
• config.json — mlx-whisper "dims" config with a quantization block ({"group_size": 64, "bits": 8})

Notes

• Loads onto the Metal GPU with no CoreML/ANE specialization step, so startup is seconds rather than the minutes a CoreML Whisper spends specializing on first launch / after every app update.
• int8 is effectively lossless vs fp16 (measured 2.6% WER on Russian FLEURS, identical to the fp16 MLX turbo); the int4 build regresses ~1.5 WER points.