divydeep/granite-speech-4.1-2b-mlx-4bit
Automatic Speech Recognition • 1B • Updated • 36
Granite Speech 4.1 2B MLX Quantizations
Apple Silicon MLX conversions of IBM Granite Speech 4.1 2B, from memory-oriented 4-bit to quality-oriented 8-bit. Use with mlx-audio.
Note 4-bit: smallest option; 2.46 GB safetensors. Preserves sample words and accents with minor French punctuation differences.
divydeep/granite-speech-4.1-2b-mlx-5bit
Automatic Speech Recognition • 1B • Updated • 49Note 5-bit: balanced MLX affine conversion with protected acoustic stack and sensitive projection layers.
divydeep/granite-speech-4.1-2b-mlx-6bit
Automatic Speech Recognition • 1B • Updated • 98Note 6-bit: quality-oriented MLX affine conversion with BF16 acoustic stack, embedding, and output head.
divydeep/granite-speech-4.1-2b-mlx-8bit
Automatic Speech Recognition • 1B • Updated • 51Note 8-bit: closest to the BF16 reference; exact match on the bundled multilingual sample smoke test.