redseaplume/Voxtral-4B-TTS-2603-MLX-4bit

Voxtral 4B TTS, MLX 4-bit Quantized

4-bit quantized weights for Voxtral 4B TTS on Apple Silicon via MLX.

Backbone: 4-bit (group_size=64), ~2.6 GB (down from ~6.8 GB BF16) Acoustic transformer: BF16 (unchanged) Vocoder: BF16, pre-processed (weight-norm reconstructed, conv weights transposed, codebook precomputed)

Total file size: 3.4 GB (vs 7.5 GB BF16)

Usage

Requires the code from redseaplume/Voxtral-4B-TTS-2603-MLX. Point model_path at this repo.

What's in here

• consolidated.safetensors: all three components in one file
• params.json: model config
• tekken.json: tokenizer
• voice_embedding/: 20 pre-computed voice embeddings (.pt and .npz)

Notes

• Only the backbone is quantized. Acoustic transformer and vocoder stay BF16.
• Generation output differs slightly from BF16 (quantization is lossy). Frame counts may vary.