onyx-whisper-tiny (mixed 4-bit MLX)
MLX quantization of openai/whisper-tiny (Apache-2.0) for the Onyx on-device speech-to-text engine (OnyxSTT). Mixed precision: the AUDIO ENCODER is 4-bit (group_size 64); the DECODER + embeddings + convs stay fp16. ~66 MB (from ~151 MB fp32). Whispers autoregressive decoder is quant-sensitive (full 4-bit mishears numbers/codes), so only the robust encoder is 4-bit. Verified WER-neutral vs fp32 on clips with numbers/names/codes. Inherits the base Apache-2.0 license.
- Downloads last month
- 57
Model size
31.8M params
Tensor type
F16
路
F32 路
U32 路
Hardware compatibility
Log In to add your hardware
Quantized
Model tree for wabibito/onyx-whisper-tiny
Base model
openai/whisper-tiny