Ant Group's Ming-3.0 Tops Open-Source Voice Charts With Five Speech Models
InclusionAI drops five Ming-3.0 speech models under MIT, led by a 7.9B MoE TTS that tops open-weight controlled voice leaderboards.
- InclusionAI open-sourced five Ming-3.0 speech models under MIT across TTS, audio, and ASR families.
- Ming-3.0-tts-tiny is a 7.9B MoE with 1.3B active parameters built on Ling V3.
- Ranks #1 open-weight on Artificial Analysis Controlled Voice (1069 Elo) and Pronunciation Robustness (93.2%).
- Supports 7 languages, Sichuanese/Cantonese dialects, 14 paralinguistic tags, voice design, and podcast dialogue.
- Day-zero serving via vLLM-Omni and SGLang-Omni through OpenAI-compatible /v1/audio/speech endpoints.
- Weakest at Korean CER (6.00%) and open-ended voice design versus Qwen3-TTS.
Ant Group’s inclusionAI team has released five Ming-3.0 models on Hugging Face and ModelScope. The MIT-licensed collection covers text-to-speech, audio understanding, tool calling, spoken dialogue, and automatic speech recognition.
The flagship TTS model card describes a 7.9-billion-parameter mixture-of-experts model that activates 1.3 billion parameters for each token. This architecture routes work through a subset of the model, reducing active computation. Deployments must still account for the full checkpoint’s storage, memory, and serving overhead.
Five checkpoints, three jobs
| Model | Size | Primary uses |
|---|---|---|
| Ming-3.0-tts-0.5B | 0.5B dense | Voice design and paralinguistic control |
| Ming-3.0-tts-tiny | 7.9B total, 1.3B active | Voice cloning, long-form narration, and multi-speaker dialogue |
| Ming-3.0-audio-flash | 126B total, 6.5B active | Audio understanding, reasoning, dialogue, and tool calling |
| Ming-3.0-audio-tiny | 9.3B total, 2.7B active | Lower-cost audio dialogue |
| Ming-3.0-asr-tiny | 8.6B total, 2B active | Accented, dialectal, and context-aware transcription |
All five checkpoints use the MIT license, which permits commercial use and modification subject to its notice and license conditions. inclusionAI has not announced a hosted paid API for the release, leaving users responsible for inference infrastructure and operations.
TTS controls voice, style, and delivery
The TTS family supports zero-shot voice cloning from a short reference clip, speaker design from plain-language descriptions, and controls for emotion, speed, pitch, and volume. Language coverage includes Chinese, English, German, Spanish, French, Korean, and Japanese, with additional controls for Sichuanese and Cantonese.
Fourteen inline tags add breaths, coughs, sighs, laughter, hesitation, surprise, and questioning or confirmation sounds. A podcast task generates multi-speaker dialogue from a script and one reference clip for each speaker.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
