Ant Group has released a new family of open speech models under its inclusionAI label. According to TechNode, which cited a report by IT Home, Ant made five Ming-3.0 models available on October 20, 2026, spanning speech synthesis, audio understanding and conversation, and speech recognition, and said they already power services including its Afu health app, Alipay search and internal office tools. The GitHub repository for the text-to-speech models, inclusionAI/Ming-3.0-tts, was created on October 19 and is released under the MIT licence.
The speech synthesis side has two models that share one generation framework, in which text and task-specific control tokens condition a language model with a diffusion-transformer patch head and an audio tokenizer. Ming-3.0-tts-0.5B is a dense model built on Qwen2.5-0.5B. Ming-3.0-tts-tiny is a mixture-of-experts model built on Ant's Ling V3 backbone, with 7.9 billion parameters in total and about 1.3 billion active per token; despite its name, its checkpoint is the larger download, at about 18 GB on Hugging Face.
Both models can clone a voice from a short reference clip, design a new voice from a written description, or use one of more than 100 built-in Chinese voices and five English ones. They accept natural-language instructions about delivery, explicit emotion, speed, pitch and volume settings, and 14 inline tags for sounds such as laughter, sighs, coughs and hesitation. They speak Chinese, English, German, Spanish, French, Korean and Japanese, plus Sichuanese and Cantonese, and output 44.1 kHz audio. Only the tiny model can generate a multi-speaker dialogue, such as a podcast, from a script and speaker references.
The README cites rankings on Artificial Analysis's text-to-speech leaderboards from October 2026: first among open-weight models on the Controlled Voice arena, with an Elo of 1069 and 13th overall; first overall on Pronunciation Robustness, at 93.2%; and third among open-weight models on Provider Voice. Ant's own evaluation tables in the repository cover further benchmarks. The other three models in the family, according to coverage of the launch, are two audio-understanding models, including a large mixture-of-experts model called audio-flash, and a speech recognition model.
In practice, the release is aimed at developers who can run research-grade code. The Hugging Face pages hold only checkpoints, and inference needs the GitHub repository; Ant says it tested on NVIDIA H800 and H20 data-centre GPUs with CUDA 12.4, the tiny model needs its own environment because its pinned dependencies differ, and accelerated inference through vLLM-Omni is listed as done. A technical report is still on the project's roadmap. Three days after the repository opened, it had about 175 stars on GitHub.
For builders of voice agents and audio products, a permissively licensed model that combines voice cloning, voice design from text and fine control over paralinguistic sounds is a useful option alongside proprietary services. Voice cloning also carries obvious misuse risks, and the repository's documentation does not describe specific safeguards beyond the licence.
Most performance claims come from Ant. The Artificial Analysis rankings are an outside measurement but were reported here through the project's README, rankings change as new models are added, and no technical report has yet been published to explain training data or methods.