Qwen3-tts-0.6b-gguf-for-koboldcpp

Hi, welcome to use,

1. The audio tokenizer gguf file in iq4xs, even i say so, only little smaller and faster than q8_0;
2. This repo's qwen3-tts-0.6b-q8_0.gguf is smaller. iq4xs quant hurts the voice clone ablity, but maybe we don't really care that little lose.
3. Updated new quanted iq4_xs tokenizer, which downcasted some fp16 tensors to q8_0.(211M)
4. Updated iq3s tokenizer, yes, it's working well and faster
5. Updated new quanted q5k, which can handle longer transcript than q4 as q8.
6. Updated MXFP4 which is faster Now you only need 703 MB RAM in total to run this model!

Note:

1. Only use 4bit base for single short paragraph or realtime llm output transcript, it will not work even for a short article !!! (maybe i should delete them.)
2. All WAvtokenizer functioning well.
3. Recommand use q5K Base + MXFP4 Tokenizer for robust, speed, and save VRAM, which is 869 M in total.