Ming-3.0-tts-tiny

Ming-3.0-tts-tiny

A multilingual text-to-speech model from inclusionAI that clones a voice from a short clip, designs voices from a description, and controls emotion, speed and dialects. Inference needs its GitHub project.

Made by inclusionAI on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Ming-3.0-tts-tiny is the small mixture-of-experts checkpoint of Ming-3.0-tts, a speech synthesis model from inclusionAI. Its Ling V3 backbone has 7.9B parameters in total, with about 1.3B active per token.

The repository holds the sharded weights, supporting model code, an audio tokenizer, a speaker encoder and a text tokenizer. It supports seven languages plus Sichuanese and Cantonese dialects.

Who it suits: Developers and audio teams who want expressive, controllable speech in Chinese, English and several other languages, and who are comfortable setting up a Python research environment with a CUDA GPU.

Industry | Who uses it | How they use it | Example
Media | Podcast and audio producers | Generate multi-speaker dialogue from a script | A two-host episode drafted from a written script
Software | App developers | Add spoken replies in several languages | A voice assistant that answers in Cantonese
Education | Language-course makers | Produce spoken examples with set emotion and speed | Slow, clear sentences for learners
Entertainment | Game and video creators | Design voices from a written description | A narrator voice described in plain words

You can skip it if you need an easy install with no GPU or an off-the-shelf app, since this is a research checkpoint with a pinned environment.

Where to use it: The model runs from its GitHub project's command-line script or Python API, on Windows, macOS or Linux with a CUDA GPU.

• Command line, through the infer_tiny.py script
• Python code, through the MingTinyTTS class
• A dedicated conda environment with Python 3.10.15
• An NVIDIA GPU, with the card recommending an H800

Use case | Who | What they get back
Voice cloning | Audio producers | Speech in a voice taken from a short reference clip
Voice design | Creators | A voice made from a written speaker description
Dialect speech | Regional-content teams | Sichuanese or Cantonese audio from text
Podcast dialogue | Podcasters | Multi-speaker audio from a script and speaker references

How to set it up: You need Python 3.10.15 in a dedicated conda environment, the pinned dependencies, and the project's GitHub code, because the Hugging Face repository does not provide a standalone inference setup.

• conda and Python 3.10.15
• PyTorch and torchaudio 2.7.1 with CUDA 11.8, and transformers 4.46.3
• A CUDA GPU, with an NVIDIA H800 recommended
• Room for about 18 GB of BF16 weights

1. Clone the Ming-3.0-tts repository from GitHub and open its folder.
2. Create the conda environment and install the tiny requirements.
3. Download the checkpoint with the hf command.

cd Ming-3.0-tts
conda create -n ming-tiny python=3.10.15 -y
conda activate ming-tiny
pip install -r requirements-tiny.txt

pip install "huggingface_hub[cli]"
hf download inclusionAI/Ming-3.0-tts-tiny --local-dir Ming-3.0-tts-tiny

How to use it: Run the infer_tiny.py script with a task, your text, a reference wav and an output path. The command below asks for Cantonese speech in the style of a reference clip and saves it to a wav file.

python infer_tiny.py \
  --model_path inclusionAI/Ming-3.0-tts-tiny \
  --task dialect \
  --text "今日天气几好，我哋食完饭一齐去海边行下啦。" \
  --dialect 广粤话 \
  --prompt_wav data/wavs/10002287-00000094.wav \
  --output_path output/tiny.wav

In Python, load the model once and call generate. The card says generate has built-in defaults for guidance scale, temperature, flow steps and sigma.

from ming_tiny import MingTinyTTS

tts = MingTinyTTS.from_pretrained("inclusionAI/Ming-3.0-tts-tiny", device="cuda:0")
tts.generate(
    task="seed",
    text="傍晚的风从山口吹过来，远处的云慢慢染成了金色。",
    prompt_wav="data/wavs/10002287-00000094.wav",
    prompt_text="在此奉劝大家别乱打美白针。",
    target_language="Chinese",          # English for English text
    output_path="outputs/seed.wav",
)

Tool | What it does
infer_tiny.py | Command-line script that turns text into a wav file
ming_tiny.MingTinyTTS | Python class that loads the model and generates speech
hf download | Fetches the checkpoint files to a local folder

Example: You give the script a Cantonese sentence about nice weather and a walk to the sea, plus a short reference recording. It writes a wav file, output/tiny.wav, with the sentence spoken in Cantonese.

For newcomers and experts: If you are new, treat this as a research tool rather than an app. The first small thing to try is the Cantonese command above after you finish the setup, then swap in your own text. The card recommends built-in voices named UK_George, US_Emma, 灵小柔 and 灵楷.

If you know the field, the Ling V3 mixture-of-experts backbone activates about 1.3B of its 7.9B parameters per token. Tiny needs its own pinned environment, differing from the 0.5B model. The card lists 14 inline tags for sounds like breath, laughter and hesitation, and a thinking_v1 prompt protocol.

Clone the GitHub project
Create the pinned conda environment
Download the checkpoint
Run infer_tiny.py or MingTinyTTS
A wav file is written

Good to know: Voice cloning copies a voice from a short clip, so use only voices you have the right to use. The Hugging Face page is for distributing the checkpoint, and the project's README holds the current dependencies and usage.

• License: MIT
• Cost to run: none named, but it expects a CUDA GPU
• Data it reaches: the reference audio you supply, plus the checkpoint and dependencies you download
• Not for: standalone use from the Hugging Face repository alone, since inference needs the GitHub project

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: transformers
• Parameters: 9B
• Weights: about 18 GB in BF16; about 4.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 9,080
• Likes: 61
• License: mit
• Languages: zh, en, de, es, fr, ko, ja
• Quantized versions: 8
• Fine-tunes: 3
• Tags: transformers, onnx, safetensors, lingv3_tts, feature-extraction, audio, text-to-speech, voice-cloning, multilingual, mixture-of-experts, speech-generation

Links

• Model: https://huggingface.co/inclusionAI/Ming-3.0-tts-tiny
• Files: https://huggingface.co/inclusionAI/Ming-3.0-tts-tiny/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download inclusionAI/Ming-3.0-tts-tiny

Then follow the usage in its model card. Nothing in this kit runs by itself.