shraey/dots-tts-mlx

Card: honest peak-memory note (~6-13GB, activation-bound)

Card: document MeanFlow (mf-int4/mf-int8) variants + honest memory note

Add MeanFlow int8 weights (mf-int8): distilled NFE-4 few-step decoder, ~2x faster

Add MeanFlow int4 weights (mf-int4): distilled NFE-4 few-step decoder, ~2x faster

Upload core.safetensors

int8: small files (vocoder/speaker/config/tokenizer)

model card: list all 24 languages (same as upstream); clarify the 5-lang check is quant-validation only

add model card (sizes + relative no-regression quality + attribution)

add int4 MLX weights (LLM-only quantized)

initial commit