Encoder-only INT8 variant of nvidia/parakeet-tdt-0.6b-v3 converted for MLX / Apple Silicon inference via mlx-audio-swift.

Scales/biases sidecars are emitted at <path>.scales / <path>.biases (sibling keys, not suffixes of .weight) and cast to FP16 — this matches the mlx-audio-swift loader's expectation.

Motivation: the Conformer encoder dominates matmul bandwidth; its matrices are large enough that INT8 weight quantization amortizes dequant overhead and reduces memory pressure. The decoder's many small matmuls do not amortize dequant cost, so keeping them FP16 avoids regressions seen with whole-model INT4.

Load via the MLXAudioSTT fork's Parakeet loader — the quantization section in config.json is auto-detected; only layers that have a .scales sibling in the safetensors will be quantized at load time.