FireRedAudio · T8star-Aix distribution
This repository provides the pinned FireRedAudio model snapshot used by both:
The source project and model are maintained by FireRedTeam/FireRedAudio. This mirror provides pinned, resumable T8star-Aix revisions. Every revision keeps the same top-level layout and includes its own integrity and quantization metadata. Review the upstream license, model documentation, and limitations before use.
Profiles
| Revision | Download | Status | Capabilities |
|---|---|---|---|
bf16-slim-v1 |
about 21.19 GiB | stable, lossless | Full stack, original BF16 main weights and inference-only RedAE decoder |
int8-wo-safe-v1 |
about 14.76 GiB | stable, recommended | Full stack, TorchAO A16W8 Qwen blocks; audio modules stay in released precision |
int8-wo-extended-v1 |
about 13.58 GiB | experimental | Full stack, extended A16W8 including audio blocks |
int8-convrot-experimental-v1 |
about 14.76 GiB | experimental | Full stack, Comfy-Kitchen ConvRot A8W8 |
All revisions above include RedAE_decoder/model.safetensors. GGUF is not provided because existing GGUF runtimes cannot execute the complete audio encoder, continuous latent, DiT, and RedAE graph.
Required layout
The same directory can be used by Desktop and ComfyUI:
FireRedAudio/
├─ FireRedAudio/
│ ├─ config.json
│ ├─ model-00001-of-*.safetensors
│ ├─ ...
│ ├─ model.safetensors.index.json
│ ├─ processor_config.json
│ ├─ tokenizer.json
│ └─ tokenizer_config.json
└─ RedAE_decoder/
└─ model.safetensors
Desktop installation
The desktop application has a built-in downloader. Choose Hugging Face or Auto, select Lite/Full, and click 开始/继续下载. Interrupted downloads resume using Hugging Face Hub metadata.
For a manual download:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="t8star/Firered-Audio-Comfy",
revision="int8-wo-safe-v1",
local_dir=r"D:\FireRedAudio-Models",
)
Then select D:\FireRedAudio-Models as the model root in the desktop application.
ComfyUI installation
Install the node first, then run its downloader:
cd ComfyUI\custom_nodes\comfyui-fireredaudio-T8
python scripts\setup_runtime.py
python scripts\download_models.py --target "D:\ComfyUI\models\TTS\FireRedAudio" --profile full --source huggingface
Or use snapshot_download directly with this destination:
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="t8star/Firered-Audio-Comfy",
revision="int8-wo-safe-v1",
local_dir=r"D:\ComfyUI\models\TTS\FireRedAudio",
)
Restart ComfyUI and select FireRedAudio in the T8 FireRedAudio Model Loader node. Do not place these dependencies into ComfyUI's Python environment; the node uses its own pinned worker runtime to keep Transformers 5.8 isolated from other nodes.
Integrity and reproducibility
The integrations pin a specific immutable Hub revision. Every profile contains fireredaudio-model.json with its exact file sizes and SHA-256 values; INT8 profiles also contain FireRedAudio/fireredaudio_quantization.json. Desktop and ComfyUI perform a fast structural/size validation after selection and can optionally perform a full SHA-256 validation.
Responsible use
Only clone or edit voices with the speaker's authorization and in compliance with applicable law and platform policies. Do not use generated audio for impersonation, fraud, harassment, infringement, or deceptive attribution. Generated output may contain errors or artifacts and requires human review before publication.