Ming-3.0 speech models
Ant Group open-sourced the Ming-3.0 speech family, five models for speech synthesis, audio understanding and conversation, and speech recognition, with the TTS models under the MIT license.
The family adds voice cloning, voice design from plain-text descriptions, and control over emotion, emphasis and pacing. The Ming-3.0-tts models come in two sizes: a 0.5b dense model and a tiny mixture-of-experts model with 7.9 billion total and 1.3 billion active parameters. The audio models handle transcription, summarization and analysis of long recordings while keeping speaker, emotion and background sound information.
- Date
- Tuesday 20 October 2026
- Lab
- Ant Group
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Ming-3.0-tts-tiny parameters | 7.9B total, 1.3B active mixture-of-experts on Ling V3 architecture | company |
| Pronunciation Robustness, correctly pronounced highlighted spans | 93.2% first overall on the Artificial Analysis leaderboard as of October 2026, per the repository | company |
Details on the TTS models come from Pandaily's reading of the repository. The other three models and their licenses were not described in detail.
Sources
- pandaily.com/ant-group-inclusionai-ming-3-0-tts-open-source-0-5b-tiny-moe-voice-control
- technode.com/2026/10/20/ant-group-open-sources-ming-3-0-speech-models-with-voice-design-an
This record was checked against its sources on 20 October 2026. How we check