Ant Group open-sources Ming 3.0 TTS

Ming 3.0 TTS is an open-source text-to-speech model family from Ant Group that turns text into realistic human speech, steered entirely by everyday conversational prompts.

Beyond cloning voices from short reference audio clips, its standout trick is free-form voice design. You can describe the tone, texture, or vibe you want in plain words—like asking for a voice that evokes an old radio crackling softly during a storm, or a dramatic narrator reading an eerie audiobook.

For granular editing, the system lets you insert 14 inline paralinguistic tags into transcripts, triggering real-sounding sighs, chuckles, and laughter, alongside direct controls for emotion, pitch, speed, and volume.

What it can do

Ming 3.0 TTS handles English, Chinese, German, Spanish, French, Korean, and Japanese, as well as Chinese dialects like Cantonese and Sichuanese.

The release splits into two distinct variants. The lightweight model is a 0.5B dense architecture, weighing in at around 3GB. The larger tiny model is an MoE build based on Ling V3, packing 7.9B total parameters with 1.3B active, taking up about 18GB.

The larger variant uniquely supports generating multi-speaker dialogues and podcasts directly from a script. Keep in mind that features operate in dedicated task modes rather than arbitrary mixes.

On the October 2026 Artificial Analysis TTS leaderboards, Ming 3.0 TTS took 1st place among open-weight models on Controlled Voice, and ranked 1st overall on Pronunciation Robustness. It scored 96.5% on text normalization and 90.6% on sequence fidelity, while placing 3rd among open-weight options on Provider Voice.

Community testers noted that while its raw voice quality is only slightly ahead of Fish Audio's Drama 3 rather than a massive generational leap, its pronunciation consistency is remarkably solid.

How to run it

The roughly 3GB lightweight model runs comfortably on consumer GPUs, falling back cleanly to PyTorch SDPA without requiring flash-attn. Ant Group tested it on NVIDIA H800 and H20 cards.

The 18GB MoE variant is far heavier, requiring its own isolated Python 3.10 and PyTorch 2.7.1 setup, with NVIDIA H800 recommended for smooth inference.

Weights and inference scripts are live on GitHub, Hugging Face, and ModelScope. Crucially, both models are released under the permissive MIT license, meaning full freedom for modification and commercial deployment.

Why it matters

If you build games, audiobooks, or synthetic podcasts, Ming 3.0 TTS is worth testing immediately. The 3GB version puts high-fidelity local TTS within reach of everyday consumer hardware, and the MIT license removes the licensing headaches common in this category.

Against rivals like Fish Audio Drama 3, it wins primarily on dialect coverage and rock-solid pronunciation accuracy. We think the natural-language steering interface makes voice tuning significantly faster than sliding rigid acoustic parameters, even if the larger 18GB model remains out of reach for modest home setups.