Ant Group's inclusionAI team has open-sourced Ming-3.0-tts, a speech-generation family that aims to put fine-grained voice control into models small enough for local deployment. The code, a task cookbook and evaluation tables are on GitHub, with weights on Hugging Face and ModelScope, all under the MIT license.
The release has two models. Ming-3.0-tts-0.5b is a dense model of about half a billion parameters. Ming-3.0-tts-tiny is a mixture-of-experts model built on Ant's Ling V3 architecture, with 7.9 billion total parameters and 1.3 billion active per token. Both share one generation framework: text and task-specific control tokens condition a language model that drives a DiT patch head and an audio tokenizer.
The feature list is broad. Users can clone a reference voice, design a new voice from a plain-text description, or pick from more than 100 built-in Chinese voices and five English ones. Delivery can be steered with natural-language instructions, and emotion, speed, pitch and volume can be adjusted directly. Fourteen inline tags insert laughter, sighs and other vocal sounds in Chinese or English. The models cover Chinese, English, German, Spanish, French, Korean and Japanese, plus Sichuanese and Cantonese. The tiny model adds multi-speaker dialogue generated from a script and speaker references, aimed at long-form podcast-style audio.
The team notes these capabilities run as separate task modes rather than arbitrary combinations.
According to the repository, Ming-3.0-tts ranked first among open-weights models on the Artificial Analysis Controlled Voice leaderboard and first overall on Pronunciation Robustness, with 93.2% of highlighted spans judged correctly pronounced, as of October 2026. The team cautions that these rankings shift as new models are added. On open benchmarks, it reports state-of-the-art results among listed open-source models on all six NVBench single-tag paralinguistic metrics, on Cantonese base-set error rate for the 0.5b model, and on German and Spanish word error rate for the tiny model.
The README is also candid about weak spots. Speaker similarity in the newly added languages mostly misses the top three, pitch control is the least accurate attribute, the tiny model's Korean error rate trails the smaller model by a wide margin, and podcast audio quality still sits below an earlier Ming-omni-tts release. Emotion accuracy drops when the text itself carries no emotional cue.
The roadmap lists vLLM-Omni accelerated inference and a technical report as next steps. The project asks users not to clone real people's voices without consent and to label synthetic speech where it could be mistaken for a real recording.
SEE ALSO: Ant Group Open-Sources Ling-3.1-flash Weights Under MIT · VUI Labs' Luna-TTS Tops the Global TTS Arena