The inclusionAI team from Ant Group has released Ming-3.0-tts, a family of speech generation models, to the public. The goal is to provide detailed control over the voice in models that are compact enough for local deployment. The code, task guide, and evaluation tables are available on GitHub, while the weights are hosted on Hugging Face and ModelScope; all materials are distributed under the MIT license.
The release includes two models. Ming-3.0-tts-0.5b is a dense model with approximately half a billion parameters. Ming-3.0-tts-tiny is a mixture-of-experts model built on Ant's Ling V3 architecture, which has a total of 7.9 billion parameters, with 1.3 billion active per token. Both models utilize a unified generative structure: a language model, governed by task-specific control tokens and text, influences the DiT patch-head and audio tokenizer.
The capabilities of these models are extensive. Users can clone a reference voice, create a new voice based on a textual description, or choose from over 100 built-in Chinese voices and five English ones. Speech delivery can be regulated using natural language instructions, and parameters such as emotion, speed, pitch, and volume can be adjusted directly. Fourteen built-in tags allow for the insertion of laughter, sighs, and other vocal sounds in both Chinese and English. The models support Chinese, English, German, Spanish, French, Korean, and Japanese, as well as Sichuanese and Cantonese.
The tiny model additionally supports multi-speaker dialogues generated from scripts and voice samples, which is geared towards creating audio in the style of long podcasts. The team notes that these capabilities function as separate task modes rather than arbitrary combinations.
According to the repository, Ming-3.0-tts ranked first among open-weight models in the Artificial Analysis Controlled Voice leaderboard and achieved first place overall in the Pronunciation Robustness Score, reaching 93.2% correctly pronounced highlighted segments as of October 2026. However, the team warns that these rankings change as new models are added. On open benchmarks, the model demonstrates state-of-the-art results among listed open-source models across all six NVBench paralinguistic metrics with one tag, error frequency on the base Cantonese dataset for the 0.5b model, and word error frequency in German and Spanish for the tiny model.
The README file also honestly lists weaknesses. Voice similarity in recently added languages generally does not make the top three leaders, pitch control is the least accurate attribute, the error rate of the tiny Korean model lags significantly behind the smaller model, and the audio quality for podcasts still falls short of the earlier Ming-omni-tts release. The roadmap indicates accelerated vLLM-Omni inference and a technical report as next steps. The project urges users not to clone real people's voices without consent and to label synthesized speech where it might be mistaken for a real recording.