Ming-3.0: Ant Group Open-Sources MIT Voice Design TTS and ASR

Introduction

Here's a voice model you can talk into being. On October 19, Ant Group's inclusionAI team open-sourced Ming-3.0, five speech models in three families: text-to-speech, audio understanding, and speech recognition. Weights and code are all MIT-licensed on Hugging Face and ModelScope. My favorite is the smallest TTS model. Ming-3.0-tts-0.5B is a 0.5B-parameter TTS model that can design a voice from a written description, clone one from a short clip, or take stage directions in plain English, and it outputs 44.1 kHz audio. You can try the family today in free ZeroGPU Spaces, run the 0.5B on a Mac through an MLX port, or clone the GitHub repo for your own NVIDIA GPU.

What Shipped

The five models

According to the

View Post on X

and the model cards:

Ming-3.0-tts-0.5B: a dense model on a Qwen2.5-0.5B backbone. It does every TTS task except podcasts.
Ming-3.0-tts-tiny: a mixture-of-experts model on Ant's Ling V3 backbone, 7.9B total parameters with about 1.3B active per token. It adds multi-speaker podcast dialogue from a script.
Ming-3.0-audio-flash: a 126B-parameter audio-language model (6.5B active). It listens to audio and answers in text, or in speech with an optional "Talker" voice. It covers transcription, long recordings, music, ambient sound, and spoken questions.
Ming-3.0-audio-tiny: the compact version, 9.3B total with 2.7B active.
Ming-3.0-asr-tiny: Chinese and English speech recognition on the Ling-tiny-v3 backbone, with dialect support and "Context ASR". Feed it hotwords or earlier dialogue to get names and jargon right. The launch post claims a 4.41% average error rate across four ASR categories.

What the TTS models can do

Three ways to pick a voice: clone a reference clip, design one from a text description, or use a built-in voice. There are 100+ Chinese voices and five English ones: US_Emma, US_James, US_Stella, UK_George, and UK_Charlotte.
Freeform direction: one natural-language instruction for tone, pacing, and emphasis.
14 inline sound tags, like [Laughter], [Sigh], [Breathing], [Cough], and [Uhm], placed exactly where you want them in Chinese or English text.
Seven languages (Chinese, English, German, Spanish, French, Korean, Japanese) plus Sichuanese and Cantonese.
Emotion (seven settings), plus speed, pitch, and volume.

One catch: these are separate task modes. You can't stack voice design, emotion, and a dialect in one call.

The rankings

The repo and launch post cite the independent Artificial Analysis TTS leaderboards as captured in October 2026. There, the 0.5B scores 1069 Elo on Controlled Voice, which makes it 1st among open-weights models and 13th overall. In that test every model clones the same eight English voices. It's tied for 1st overall on Pronunciation Robustness at 93.2%, level with Fish Audio Drama 3 Preview, and 3rd among open weights on Provider Voice at 1188 Elo. Arena Elo shifts as new models land, so treat these as a snapshot. The other benchmark tables in the README are the team's own runs.

Try It

The quickest taste is in the browser: mrfakename's ZeroGPU Spaces for tts-tiny, asr-tiny, and audio-tiny. The TTS Space has tabs for voice clone, instruction, voice design, built-in voices, emotion and pace, and dialect.

To run the 0.5B locally on an NVIDIA GPU:

git clone https://github.com/inclusionAI/Ming-3.0-tts.git
cd Ming-3.0-tts
pip install -r requirements.txt
hf download inclusionAI/Ming-3.0-tts-0.5b --local-dir Ming-3.0-tts-0.5b

Then design a voice in Python, with no reference audio needed:

from ming_tts import MingTTS

model = MingTTS("inclusionAI/Ming-3.0-tts-0.5b")
model.synthesize(
    task="VOICE_DESIGN",
    text="The glacier moves slowly, quietly recording the climate.",
    controls={"style": "A mature, steady male documentary narrator with clear diction, an unhurried pace and a calm, thoughtful tone."},
    use_zero_spk_emb=True,
    output_path="output/design.wav",
)

Or use one of the built-in English voices:

model.synthesize(
    task="IP",
    text="The garden will remain open until sunset, provided the weather stays clear.",
    controls={"ip": "US_James"},
    use_zero_spk_emb=True,
    output_path="output/james.wav",
)

A heads-up on hardware: the team has only tested the 0.5B on NVIDIA H800-80GB and H20-96G cards. flash-attn is optional, and inference falls back to PyTorch SDPA. Tiny needs its own environment (Python 3.10.15, PyTorch 2.7.1 with CUDA 11.8), and the team recommends an H800. Mac folks can use the community mlx-community/Ming-3.0-tts-0.5B-8bit port for mlx-audio. It's 1.5 GB on disk and peaks at 2.2 GB of memory on an M4 Pro. For now it needs the branch from mlx-audio PR #1047, which isn't merged yet. Older mlx-audio releases load the model but say the wrong words.

Here are a few voice-design starters to paste into the style field or the Space's "Voice description" box:

A warm young adult woman with a soft, slightly husky timbre, an unhurried late-night radio pace and a gentle, reassuring tone.
Gender: male. Age: middle-aged. Pitch: low. Speed: slow and deliberate. Timbre: rich and resonant. Emotion: calm confidence.
A wise old ship captain telling a ghost story by the fire, gravelly voice, dramatic pauses, quiet menace under a friendly smile.
A bright, crisp product-demo host in her twenties, upbeat and precise, smiling through every sentence without sounding salesy.

My tip: put only the spoken words in text and keep every direction in style or instruction. Every voice-design example in the repo and the Space is written in Mandarin, and the voice-design benchmark is Chinese, so if an English description drifts, try writing it in Chinese. For quick changes to delivery, the FREEFORM task with an English instruction like "Speak in a warm, intimate storytelling voice, unhurried" is the easier route.

Why It Matters for Makers

A 0.5B open model that tops the open-weights board on a cloned-voice arena and designs voices from a sentence is a big deal for audiobook makers, game devs, and short-video dubbers on a budget, and MIT means you can ship it. The trade-offs are real, though. There's no technical report yet (it's still an unchecked box on the roadmap). Ming-3.0 isn't on fal or Replicate as of this writing, so you're self-hosting or using community Spaces. The team's own limitations list flags weaker speaker similarity in the five newer languages, shakier results with many tags in one line, and lower role-play voice-design scores. And the star of the audio family, Audio Flash at 126B, is a data-center model: its reference serving profile spans eight 80 GB H800s. As always, the team asks you not to clone real people without consent, and to label synthetic speech.

Conclusion

Ming-3.0 is Ant Group's MIT-licensed speech family: two TTS models, two audio-language models, and a context-aware ASR model. The headline is a 0.5B TTS model that designs voices from text, clones from a clip, takes 14 sound tags, and speaks seven languages. Open the tts-tiny Space, describe a voice you've always wanted for a character, and hear it speak. Happy voice-making!

Original source:

https://github.com/inclusionAI/Ming-3.0-tts

Model cards: Ming-3.0-tts-0.5B · Ming-3.0-tts-tiny · Ming-3.0-audio-flash · Ming-3.0-asr-tiny · Demos: Blog & Demos · Launch thread: