Inclusion AI · Ant Group

Ming-3.0-tts

Open-weights speech generation for audio stories, short-video narration and product assistants. Voice design, cloning, director-style instructions, paralinguistic tags, seven languages and two dialects.

GitHub Hugging Face ModelScope
Ming-3.0-tts-0.5b on Artificial Analysis · October 2026 Details
1st among open-weights models Controlled Voice Elo 1069 · 13th overall 1st tied, overall Pronunciation Robustness 93.2% pronounced correctly 3rd among open-weights models Provider Voice Elo 1188 · 17th overall

01Architecture

One generation framework for both models.

Text and task-specific control tokens condition a language model with a DiT patch head and an audio tokenizer. Capabilities use separate task modes, not arbitrary combinations.

02Voice design

Describe a voice in words.

Audio stories, short-video narration and product assistants each need their own voice. Voice Design creates one from a text description; with a reference clip at hand, the model clones that voice instead.

03Director mode

Direct each line: emotion, rhythm and stress.

Tell the model in plain language how a line should be performed: how the emotion shifts, which words to stress, how the rhythm goes. Listen, then adjust the direction or the line.

04Paralinguistic tags

Laughs and sighs that show how a character feels.

A sigh before a line, or a short laugh halfway through it, tells listeners what state a character is in. Inline tags place the laugh, the sigh or the cough exactly where it belongs.

05Languages and dialects

Seven languages and two dialects.

For international products and audio content, Ming-3.0-tts speaks Chinese, English, Japanese, Korean, French, German and Spanish. Below, one story in all seven, in two cloned voices. Given a reference clip in the same dialect, it also speaks Sichuanese and Cantonese.

06Models

Two models: 0.5b and tiny.

Both models are built for efficient inference. Weigh content length, expressive needs and compute budget when choosing.

Dense

Ming-3.0-tts-0.5b

0.5B parameters, for lightweight deployment.

Backbone
Qwen2.5-0.5B
Tasks
Every task except podcast

Hugging Face ModelScope

Mixture of experts

Ming-3.0-tts-tiny

7.9B total and 1.3B active parameters, with long-form podcast generation.

Backbone
Ling V3
Tasks
Every task, including podcast

Hugging Face ModelScope

07Benchmarks

Strong quality at 0.5B parameters.

Ming-3.0-tts-0.5b on Artificial Analysis, October 2026 (as captured). Ranks change as new models are added.

Controlled Voice

Every model clones the same eight English reference voices, so naturalness, audio quality and rhythm are compared under one voice: a basis for choosing a model to dub from a reference. Ming-3.0-tts-0.5b scores 1069 Elo, 1st among open-weights models and 13th overall.

Open-weights models · Elo
  1. 1Ming-3.0-tts1069
  2. 2Fish Audio Drama 3 Preview1059
  3. 3Breeze TTS 21018
  4. 4Voxtral TTS1005
  5. 5OpenAudio S1 Mini1004
  6. 6Fish Audio S2 Pro1002

Bars start at Elo 800Artificial Analysis

Pronunciation Robustness

Whether proper nouns, abbreviations and numbers are read correctly: what English narration and announcements demand of key details. Ming-3.0-tts-0.5b scores 93.2%, tied for 1st, and is 1st on the Expanding Shorthand (96.5%) and Preserving Exact Sequences (90.6%) sub-scores.

All models · share pronounced correctly
  1. 1Ming-3.0-tts93.2%
  2. 2Fish Audio Drama 3 Preview93.2%
  3. 3Eleven v491.7%
  4. 4Eleven v4 Turbo90.1%
  5. 5Gemini 3.8 Flash TTS89.5%
  6. 6Gemini 3.1 Flash TTS88.2%
Show all 22 models
  1. 7Fish Audio S2.1 Pro (Oct 2026)88.1%
  2. 8SpaceXAI TTS87.6%
  3. 9Gemini 3.8 Flash-Lite TTS87.4%
  4. 10Eleven v385.6%
  5. 11V3 Conversational84.8%
  6. 12Qwen-Audio-3.1-TTS-Plus84.6%
  7. 13HeyGen Voice83.2%
  8. 14Realtime TTS-280.7%
  9. 15Simba 3.277.6%
  10. 16Sonic 3.674.5%
  11. 17Breeze TTS 273.6%
  12. 18Speech 2.8 HD73.6%
  13. 19Lightning V3.1 Pro (Jul 2026)72.1%
  14. 20Sonic 3.568.8%
  15. 21Luna TTS63.6%
  16. 22Gradium TTS (Aug 2026)62.1%

Bars start at 0%Artificial Analysis

Provider Voice

Each model speaks in its own representative voices, so the comparison is the overall impression in real use: a guide to voice options for product assistants and content dubbing. Ming-3.0-tts-0.5b scores 1188 Elo, 3rd among open-weights models and 17th overall.

Open-weights models · Elo
  1. 1Fish Audio Drama 3 Preview1244
  2. 2Breeze TTS 21222
  3. 3Ming-3.0-tts1188
  4. 4Fish Audio S2 Pro1114
  5. 5Step Audio EditX (Mar 2026)1100
  6. 6Voxtral TTS1089

Bars start at Elo 800Artificial Analysis

Open benchmarks

The full evaluation of both models also covers generation in Chinese, English, German and Spanish, paralinguistic control and voice design from text. Check separately whether speech is read correctly, whether the voice matches its description and whether tags trigger the intended sound, then choose the model for your workflow. Full tables are in the repository README.

Chinese, English, German and Spanish generation
Paralinguistic control and voice design

WER and PCER: lower is better. SIM and ACC: higher is better.

Next: more consistent voices across languages, finer emotion and role control, and steadier paralinguistic and long-form generation.

08Samples

All samples by task.

Each one is a single call to one task, with its text and controls as given.

Get started

Download and run.

Installation, inference examples and a cookbook for every task are in the repository.

GitHub Quick Start Cookbook 
pip install "huggingface_hub[cli]"
hf download inclusionAI/Ming-3.0-tts-0.5b \
  --local-dir Ming-3.0-tts-0.5b

# In mainland China, ModelScope is usually faster
pip install modelscope
modelscope download --model inclusionAI/Ming-3.0-tts-0.5b \
  --local_dir Ming-3.0-tts-0.5b