davidalarrea/fluid-speech-mini-mlx-8bit

Fluid Speech Mini — MLX (8-bit affine (group size 64))

Fluid Speech Mini by FluidVoice, built on NVIDIA Parakeet TDT 0.6B v3 and Moondream Parakeet Ultra, all CC BY 4.0. Weights via apiplant/fluid-speech-mini-safetensors, a format-only conversion of FluidVoice-Labs/fluid-speech-mini-coreml.

This is a format conversion of apiplant/fluid-speech-mini-safetensors at revision 063ca110e3642e9355f313e0a51030d04874de26, so it loads in mlx-audio and mlx-audio-swift as a Parakeet TDT model. It is the download used by OpenWhisper. The weights were expanded to fp16 and laid out like mlx-community/parakeet-tdt-0.6b-v3, then quantized to 8 bits with MLX. Accuracy can differ slightly from the original; see the upstream card for its numbers.

Reproduce with scripts/convert-parakeet-models.py fluid-speech-mini --bits 8 in the OpenWhisper repository.

Licensed CC BY 4.0, as the original. See NOTICE.