File size: 1,009 Bytes
f4b768a
 
 
 
9868142
f4b768a
af40c0f
 
fbf7401
f575e1a
d8b5b95
118759f
26cad53
582c46d
62c8e8b
2dba9ae
62c8e8b
114c17c
62c8e8b
73fe700
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
---
base_model:
- Qwen/Qwen3-TTS-12Hz-0.6B-Base
pipeline_tag: text-to-speech
license: apache-2.0
---
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.