inclusionAI releases Ming-3.0-tts-tiny, a 7.9B-A1.3B MoE speech synthesis model
inclusionAI/Ming-3.0-tts-tiny
inclusionAI released Ming-3.0-tts-tiny, a mixture-of-experts TTS checkpoint built on the Ling V3 backbone with 7.9B total parameters and about 1.3B active per token, under the MIT license. It supports zero-shot voice cloning, voice design from text descriptions, free-form style control, 14 inline paralinguistic tags, emotion, speed, pitch and volume control, multi-speaker podcast scripts, and Chinese, English, German, Spanish, French, Korean, Japanese plus Sichuanese and Cantonese dialects. The Hugging Face repo only distributes weights; inference needs the Ming-3.0-tts GitHub repo, and serving is documented for vLLM-Omni and SGLang-Omni via an OpenAI-compatible /v1/audio/speech endpoint.
A 7.9B-total, 1.3B-active MoE TTS checkpoint with voice cloning, dialect and paralinguistic control, plus benchmark tables that also list its own weak spots.
Source:inclusionAI Models · huggingface.co