Ming-3.0-tts

What does Ming-3.0-tts do?

The README calls Ming-3.0-tts "multilingual speech synthesis with free-form and paralinguistic control." Two models are released: a 0.5 billion dense model (Ming-3.0-tts-0.5b) and a mixture-of-experts model (Ming-3.0-tts-tiny) at 7.9 billion total and 1.3 billion active parameters, built on Ling V3. Both share one generation framework: 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. The README lists voice cloning from a reference clip, voice design from a text description, built-in Chinese and English voices, free-form natural-language delivery, 14 inline paralinguistic tags for laughter and sighs, nine supported languages and dialects (Chinese, English, German, Spanish, French, Korean, Japanese, Sichuanese, Cantonese), emotion and attribute control (speed, pitch, volume), and multi-speaker podcast dialogue (tiny only).

Key facts

• Licence: MIT on the code. The Hugging Face model cards for both weights also carry the license:mit tag.
• Family: Ming-3.0-tts-0.5b (dense, 0.5 billion parameters) and Ming-3.0-tts-tiny (MoE, 7.9 billion total, 1.3 billion active, built on Ling V3).
• Releases: Ming-3.0-tts-0.5b and Ming-3.0-tts-tiny weights and inference code, with a task cookbook and evaluation tables, released October 2026.
• Hardware: 0.5b tested on NVIDIA H800-80GB and H20-96G. For tiny, the README recommends NVIDIA H800 for inference. flash-attn is optional for 0.5b; inference falls back to PyTorch SDPA.
• Software: PyTorch and torchaudio pinned per variant (0.5b uses transformers==4.46.3; tiny needs Python 3.10.15 and transformers==4.46.3 in its own environment). vLLM-Omni accelerated inference is on the roadmap and marked completed.
• Independent benchmark position (Artificial Analysis, October 2026, by the README): first among open-weights models on Controlled Voice (Elo 1069) and first overall on Pronunciation Robustness (93.2 percent correctly pronounced). Third among open-weights on Provider Voice (Elo 1188).
• The tiny podcast task scores first in CER (2.03 percent) and second in cpSIM (0.561) and UTMOS (2.12) on the podcast benchmark reported in the README.
• Tasks: TTS, FREEFORM, VOICE_DESIGN, MULTILINGUAL, BASIC, EMOTION, DIALECT, IP (built-in voices), plus podcast on tiny.

What does it replace, and where does it fall short?

Ming-3.0-tts replaces ElevenLabs for a team that wants an open-weights TTS family under MIT, run on their own GPUs, with voice cloning, cross-language synthesis and paralinguistic tags. It appears on the open-source alternatives to ElevenLabs page.

Where it falls short: the limitations section of the README lists several gaps. Speaker similarity on four of the five new languages (not German) does not reach the top three; French SIM is 0.675 for 0.5b and 0.733 for tiny. Pitch control is the least accurate attribute, with tiny 7.30 percentage points below 0.5b. Role-play voice design ranks fourth and fifth on the InstructTTSEval-ZH leaderboard. Tiny's Korean CER (6.000 percent) ranks sixth and well behind 0.5b's 1.728 percent. Hardware is an NVIDIA H800-class GPU for the recommended setup; CPU and consumer GPUs are not addressed by the README.

How does Ming-3.0-tts run?

Clone the repository and install the pinned dependencies for the chosen variant. For 0.5b: pip install -r requirements.txt, or build the Docker image at docker/Dockerfile. For tiny: a separate conda environment with Python 3.10.15 and requirements-tiny.txt. Text normalization uses FST grammars and needs pynini; pass --no_text_normalization to skip it.

Download the weights from Hugging Face: hf download inclusionAI/Ming-3.0-tts-0.5b --local-dir Ming-3.0-tts-0.5b, or the tiny variant. Run from the command line (infer.py or infer_tiny.py), from Python with MingTTS or MingTinyTTS, or through the one-click scripts in examples/one_click/ for the tiny variant.

Who is Ming-3.0-tts for?

A team or researcher with an NVIDIA H800-class GPU who wants an open-weights TTS family under MIT, with cross-language synthesis, voice cloning and paralinguistic tags. For a smaller TTS run under more modest hardware, see other open projects in the text-to-speech category. For a hosted product, ElevenLabs is the paid tool.

What limits does the README state?

• Speaker similarity. Tiny ranks third on German SIM (0.791), but neither model reaches the top three in the other four new languages.
• Several events in one line. NVBench Multi works, but trails Single; stability with many tags per sentence is still improving.
• Emotion from delivery alone. When the words carry no emotional signal, accuracy drops well below the Text-Related half.
• Pitch on tiny. Pitch is the least accurate attribute for both models. Tiny ranks 2nd at 78.70 percent, 7.30 percentage points below 0.5b.
• Role-play voice design. RP is the lowest voice-design sub-score, ranking 4th for tiny and 5th for 0.5b.
• Korean on tiny. Tiny's Korean CER (6.000 percent) ranks 6th, well behind 0.5b's 1.728 percent (3rd).
• Podcast audio quality. Tiny's UTMOS of 2.12 ranks 2nd, below ZipVoice-Dia's 2.24 and the historical Ming-omni-tts-0.5B score of 2.25.

Questions people ask

Is Ming-3.0-tts open source?

Yes. The LICENSE file at the repository root is MIT, an OSI-approved licence. The Hugging Face model cards for inclusionAI/Ming-3.0-tts-0.5b and inclusionAI/Ming-3.0-tts-tiny both carry the license:mit tag on the weights. Code and weights are both under MIT.

Can I use Ming-3.0-tts commercially?

The MIT licence on code and weights allows commercial use, modification and redistribution. The README adds a usage disclaimer: voice cloning and built-in character voices are "provided for research and legitimate creative use. Do not use them to impersonate a real person without consent, to deceive, or for any illegal purpose. Label synthetic speech as synthetic where it could be mistaken for a real recording."

What hardware does Ming-3.0-tts need?

The 0.5b model was tested on NVIDIA H800-80GB and H20-96G; flash-attn is optional, and inference falls back to PyTorch SDPA. For the tiny model, the README recommends NVIDIA H800 for inference. Both variants use transformers==4.46.3; the tiny environment also pins PyTorch and torchaudio to 2.7.1+cu118. The README does not state consumer GPU or CPU performance.

How does Ming-3.0-tts compare with ElevenLabs?

ElevenLabs is a hosted product with its own API. Ming-3.0-tts is an open-weights model family under MIT, run on your own hardware. On the Artificial Analysis Controlled Voice leaderboard (October 2026), Ming-3.0-tts ranked first among open-weights models (Elo 1069) and thirteenth overall; on Pronunciation Robustness it ranked first overall (93.2 percent). ElevenLabs' own scores on those leaderboards are not stated in the README. The open-source alternatives to ElevenLabs page sets Ming-3.0-tts beside other TTS projects.

What does the podcast task do?

Only the tiny variant has the podcast task. The README says it generates multi-speaker dialogue from a script and a reference clip per speaker; speaker turns are marked with speaker_1: and speaker_2: at the start of each line. On the podcast benchmark reported in the README, Ming-3.0-tts-tiny ranks first in CER (2.03 percent) and second in cpSIM (0.561) and UTMOS (2.12), with ZipVoice-Dia scoring 2.24 on UTMOS.

Sources

• Ming-3.0-tts README and LICENSE: github.com/inclusionAI/Ming-3.0-tts, read 2026-11-02.
• Ming-3.0-tts-0.5b model card: huggingface.co/inclusionAI/Ming-3.0-tts-0.5b, read 2026-11-02.
• Ming-3.0-tts-tiny model card: huggingface.co/inclusionAI/Ming-3.0-tts-tiny, read 2026-11-02.