inclusionAI open-sourced Ming-3.0-tts-0.5B, a multilingual TTS model with voice cloning
inclusionAI released Ming-3.0-tts-0.5B, a 0.5B-parameter multilingual text-to-speech model built on Qwen2.5-0.5B, on Hugging Face under the MIT license. It supports zero-shot voice cloning, voice design, and free-form delivery control in Chinese, English, German, Spanish, French, Korean, and Japanese, plus Sichuanese and Cantonese dialects. The checkpoint requires the Ming-3.0-tts GitHub repository for inference.
China context
- Original name
- 蚂蚁集团
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can download the MIT-licensed checkpoint from Hugging Face and integrate it via the Ming-3.0-tts GitHub repository, enabling multilingual TTS with voice cloning and free-form control without relying on Chinese cloud APIs.
- For investors
- The release of a compact, MIT-licensed multilingual TTS model by Ant Group's inclusionAI signals increased competition in open-source speech AI, potentially pressuring commercial TTS API pricing and accelerating adoption of voice-enabled applications.
inclusionAI (Ant Group) released Ming-3.0-tts-0.5B, a dense 0.5B-parameter multilingual text-to-speech model built on Qwen2.5-0.5B, on Hugging Face under the MIT license. The model supports zero-shot voice cloning from a short reference clip, voice design from natural-language speaker descriptions, free-form control of delivery and speaking style, 100+ built-in Chinese voices and five built-in English voices, 14 inline tags for paralinguistic sounds, explicit emotion/speed/pitch/volume control, and multilingual synthesis including Sichuanese and Cantonese dialects. The checkpoint includes the backbone, DiT head, audio tokenizer, speaker encoder, and text tokenizer, but inference requires cloning the Ming-3.0-tts GitHub repository; the Hugging Face repository is for checkpoint distribution only. The model card includes benchmark comparisons on Seed-TTS test sets, with Ming-3.0-tts-0.5B achieving a Chinese WER of 0.920% and English WER of 1.320% (company-reported).
Ming-3.0-tts-0.5B is a dense 0.5B-parameter TTS model built on the Qwen2.5-0.5B backbone, integrating a DiT head, audio tokenizer, speaker encoder, and text tokenizer. It supports zero-shot voice cloning, voice design from natural-language descriptions, free-form delivery control, 14 inline paralinguistic tags, explicit emotion/speed/pitch/volume control, and multilingual synthesis across seven languages plus Sichuanese and Cantonese dialects. Inference requires the Ming-3.0-tts GitHub repository; the Hugging Face repo is for checkpoint distribution only. The model card reports Seed-TTS benchmark results: Chinese WER 0.920%, English WER 1.320% (company-reported).
Developers outside China can now use a MIT-licensed, 0.5B-parameter multilingual TTS model with voice cloning and free-form control, reducing the cost and complexity of building custom speech applications compared to larger proprietary TTS APIs. The release of Ming-3.0-tts-0.5B by inclusionAI (Ant Group) puts pressure on other open-source TTS models like GLM-TTS and Qwen3-TTS, which show higher WER on the Seed-TTS benchmark.
For businesses needing multilingual voice synthesis, Ming-3.0-tts-0.5B offers a permissive MIT license and a compact 0.5B parameter footprint, potentially lowering inference costs and enabling on-premise deployment. The model's support for voice cloning, voice design, and free-form control could reduce the need for expensive voice actor recordings or proprietary TTS APIs.
The next verifiable signal is whether inclusionAI releases the full Ming-3.0-tts training code or larger checkpoints, and whether third-party evaluations confirm the reported Seed-TTS benchmark results. Adoption can be tracked via GitHub stars, forks, and downstream model usage on Hugging Face.