Ming-3.0-tts is an open-source text-to-speech model family from inclusionAI, Ant Group's AI team. On October 19, 2026, the inference code, model weights and full evaluation tables landed together under the inclusionAI organisation on GitHub, under the MIT licence, with weights mirrored on Hugging Face and ModelScope. The release ships two models at once: a 0.5B dense model, and a tiny variant built on Ling V3 — a mixture-of-experts design with 7.9B total parameters and 1.3B active.
The pitch is not sound quality, it is delivery
This generation of Ming moves control into the text itself: 14 paralinguistic tags can be written inline to direct laughter, sighs and other sounds beyond the words; natural-language instructions can describe the desired delivery; or a style description alone can "design" a brand-new voice. Voices come three ways — cloning a reference clip, designing from text, or picking from 100+ built-in Chinese voices and five English ones. Languages cover Chinese, English, German, Spanish, French, Korean and Japanese, plus the Sichuanese and Cantonese dialects; multi-speaker podcast-style dialogue generation is exclusive to the tiny model.
The scoreboard: first overall in pronunciation robustness
On the Artificial Analysis leaderboards published by the team (October 2026), Ming-3.0-tts ranks first among open-weights models on Controlled Voice, first overall on Pronunciation Robustness at 93.2% accuracy — also topping the Expanding Shorthand (96.5%) and Preserving Exact Sequences (90.6%) sub-metrics — and third among open-weights models on Provider Voice. Pronunciation robustness is exactly where TTS breaks in production: misread one number, abbreviation or proper noun and listeners stop trusting the whole clip, which makes this particular first place worth more than a sound-quality trophy.
Boundaries to know before you download
The two models do not share a runtime: the 0.5b was tested on H800-80GB and H20-96G hardware, while tiny requires its own reference environment (PyTorch 2.7.1 with CUDA 11.8), and the documentation insists on separate installs. Abilities are not freely composable either — each task mode accepts only its specified inputs, and combinations from other modes are rejected. The technical report and vLLM-Omni accelerated inference are still on the roadmap; what exists today is weights, code and a complete task cookbook. The official demo page hosts audio samples for every task mode, so you can listen first and decide whether the weights are worth the download.
