| This model is licensed under the Fish Audio Research License, Copyright © 39 AI, INC. All Rights Reserved. | |
| Built with Fish Audio. | |
| Changes made by ajh-code: | |
| - Quantized 60 slow-transformer gate/up projections in layers 3–32 to packed | |
| NVFP4 E2M1 weights with a native W4A16 decode path and W4A4 larger-M path. | |
| - Quantized the other 120 slow-transformer projections to native MXFP8 W8A8. | |
| - Folded calibrated English channel scales into selected normalization and | |
| packed-weight tensors. | |
| - Added a portable mixed-precision loader, Blackwell kernels, a TTS API/web | |
| interface, compact inference-only DAC buffers, and reference-encoder staging. | |
| - Kept embeddings, normalization arithmetic, output heads, the fast | |
| transformer, KV cache, sampling, and DAC arithmetic at BF16 or their original | |
| higher precision. | |
| The original model is fishaudio/s2-pro, revision | |
| 1de9996b6be38b745688de084d87a5633f714e4e. The bundled Fish Speech source is | |
| revision e5e292632cb11e7a27b2b7487f58f612bc101e13. | |