The IT Home page is stamped 2026/10/20 9:34:53. It reports Ant Ling’s announcement the same day: the Ming-3.0 speech series is open-sourced in three lines, speech synthesis, audio understanding and conversation, and speech recognition, five models in all. The names on the page are Ming-3.0-tts-0.5B, Ming-3.0-tts-tiny, Ming-3.0-audio-flash, Ming-3.0-audio-tiny, and Ming-3.0-asr-tiny. Weight links point to Hugging Face and ModelScope. The page prints no evaluation scores.

[1]
A stack of charcoal cards with a muted blue edge
A closed stack of cards, the top card painted charcoal, with a muted blue edge and warm paper edges below., AI-generated illustration, not a news photograph

The capability notes are also the company’s product description. The synthesis line says Voice Design can describe a target voice in text, can also set a voice from a reference audio clip, and includes a director mode for describing how a line should be performed. The audio-understanding line says it can transcribe, summarize, and analyze long audio directly, and that this keeps emotion, speaker identity, and setting. None of that comes with a comparison model or a sample count.

[1]

The page says the synthesis model is already used in the main chat and voice calls of the Afu app, and in Xiaohu office software. The recognition model is already used in the Afu doctor assistant, Alipay search, Yixiaobao’s live voice conversation, and a safety check. Those are statements of where it is deployed, not a public leaderboard. The page also says Ling Studio offers an interface for three of the models. How to call that interface is not in this edition.

[1]