littoralai/cohere-transcribe-mlx-8bit

8-bit MLX quantization of CohereLabs/cohere-transcribe-03-2026, via the MLX fp16 conversion at littoralai/cohere-transcribe-mlx-fp16. Quantized with mlx-audio (8-bit, group size 64, affine). No architecture, vocabulary, or training changes. See the base model card for capabilities, training details, and evaluations.

License

Apache-2.0, from the base model. The custom architecture code (modeling_cohere_asr.py, configuration_cohere_asr.py, processing_cohere_asr.py, tokenization_cohere_asr.py) is Cohere's, redistributed unchanged. Changes vs. the source: fp16 MLX conversion, then 8-bit affine quantization.

Integrity

model.safetensors sha256 aadaf8d3388975853385400c9bd8dee92a71e12a83860f35bebaac345aa8af93

Use with mlx-audio

python -m mlx_audio.stt.generate --model littoralai/cohere-transcribe-mlx-8bit \
  --audio audio.wav --output out