Ant Group open-sources Ming-3.0-tts speech models

Ant Group's inclusionAI team released Ming-3.0-tts under the MIT license: a 0.5B dense model and a 7.9B-parameter MoE tiny model with 1.3B active parameters. The family supports voice cloning, text-described voice design, 14 inline vocal tags and nine languages, and tops the Artificial Analysis open-weights voice leaderboard.

• Two models: 0.5B dense and 7.9B MoE with 1.3B active parameters
• Over 100 built-in Chinese voices and five English ones
• 14 inline tags for laughter, sighs and other vocal sounds
• Ranked first among open-weights models on Artificial Analysis