BarathwajAnandan/cohere-transcribe-03-2026-CoreML-6bit

Replace broken Cohere decoder packages with consistent 35s export

Fix Cohere manifest to match 35s CoreML artifact contract

Update cached decoder to accept pre-computed cross K/V

Update manifest with cross_kv_projector entry

Add cross-KV projector (cross-attention pre-computation optimization)

6-bit palettized CoreML: 35s window, 256 max tokens, 5s overlap

Replace with 6-bit palettized CoreML (35s window, 128 max tokens, 5s overlap)

Add GPL-3.0 license file

Update model license to GPL-3.0

Update model card with full LibriSpeech WER: 2.58%

Add full LibriSpeech test-clean WER: 2.58% on 2619/2620 samples

Upload CoreML palettize6 model (6-bit, 1.4GB, 2.72% WER on LibriSpeech test-clean)

Add benchmark: benchmarks/compute_unit_comparison.json

Add benchmark: benchmarks/quantization_variants_speed_accuracy.json

Add benchmark: benchmarks/librispeech_wer_200samples.json

Change license to GPL v3

Add model card with license, benchmarks, and usage docs

initial commit