Prepare models from ModelScope
Choose a task, model and source in Setup. The repository is filled in for you. Download to your chosen disk, then register the local model.
Five steps
- Open Setup → Models & parameters → Download models & components.
- Choose ModelScope, then a task and model shortcut. All 13 shortcuts have a verified source mapping.
- Choose your model directory, then Inspect download list. Inspection retrieves metadata without downloading weights.
- Review sizes and selected files, then explicitly choose Download selected. Defaults select only SenseVoice model.int8.onnx + tokens.txt, Whisper ggml-base.bin, or HY-MT2 Q4_K_M.
- After downloading, add the model to the local library, inspect it, prepare native files where requested, and try a short recording. A completed download does not prove inference readiness.
Repository list
Owners can differ between platforms. These are ModelScope repositories; do not assume the Hugging Face owner is interchangeable.
| Model | ModelScope repository | Default revision |
|---|---|---|
| IndexTTS 2 | IndexTeam/IndexTTS-2 | master |
| IndexTTS 2.5 | IndexTeam/IndexTTS-2.5 | master |
| VoxCPM2 | OpenBMB/VoxCPM2 | master |
| Qwen3-TTS Base | Qwen/Qwen3-TTS-12Hz-1.7B-Base | master |
| Qwen3-TTS VoiceDesign | Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign | master |
| Qwen3-TTS CustomVoice | Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice | master |
| FireRedTTS3 | FireRedTeam/FireRedTTS3 | master |
| MOSS-TTS | openmoss/MOSS-TTS | master |
| HY-MT2 GGUF | Tencent-Hunyuan/Hy-MT2-1.8B-GGUF | master |
| SenseVoice ONNX | WEAAEW/sherpa-onnx-sense-voice-zh-en-ja-ko-yue-2024-07-17Community mirror · recommended weights match upstream hashes | master |
| Fun-ASR-Nano | FunAudioLLM/Fun-ASR-Nano-2512 | master |
| Whisper GGML | cjc1887415157/whisper.cppCommunity mirror · recommended weights match upstream hashes | master |
| emotion2vec+ Base | iic/emotion2vec_plus_base | master |
A few useful distinctions
SenseVoice and Whisper use explicitly labeled community mirrors. Published SHA-256 hashes of recommended model weights match upstream; this does not endorse the uploader. SenseVoice requires ONNX rather than raw model.pt. Whisper requires whisper.cpp GGML files.
IndexTTS and VoxCPM2 may require native-file preparation. Original emotion2vec model.pt requires offline export to a labeled classifier.onnx. Full MOSS-TTS generation remains unavailable.
Changing model or source resets the repository revision instead of carrying over a previous commit. Collapsing the panel or navigating Setup retains custom settings and jobs without automatically restarting downloads. master is a mutable branch; inspect the list before each download.
Read each model’s own license and usage conditions. Catalog inclusion, hash checks and a successful download do not establish commercial permission.