FULL STORY

Ant's Ming-3.0 Open-Sourced, Tops TTS Leaderboards

Ant Group's inclusionAI open-sourced five Ming-3.0 voice models, and Ming-3.0-tts soon topped multiple Artificial Analysis TTS leaderboards, including pronunciation robustness.

2026-10-19 ~ 2026-10-20 · 2 episodes · 14 posts

Episode 1 · Ant's inclusionAI Open-Sources Five Ming-3.0 Speech Models (2026-10-19, 6 posts)

Ant Group's inclusionAI (AntLingAGI) open-sourced five Ming-3.0 voice models on Hugging Face and ModelScope on October 19–20. All use a MoE architecture, ranging from 0.5B to 126B total parameters (6.5B activated), with weights released under MIT and Day-0 inference support via vLLM-Omni.

Confirmed

  • Ming-3.0-tts family: tts-0.5B supports voice design and paralinguistic cues; tts (7.9B, 1.3B activated) excels at long-form speech generation; a post also mentions tts-tiny (9B) text-to-speech specs.
  • Ming-3.0-audio family: audio-tiny is a 9.3B compact native audio language model that takes audio + text instruction input, supporting speech recognition, spoken Q&A, audio understanding and reasoning, with optional built-in or reference voices for speech synthesis; audio-flash has 126B total / 6.5B activated parameters, supporting audio understanding + speech output.
  • The series covers controllable speech synthesis, long-recording analysis, and context-aware transcription.
  • According to @AdinaYakup and the repost by @realmrfakename, Ming-3.0-tts-0.5B tops the open-source TTS ranking on the Artificial Analysis Controlled Voice leaderboard, with voice cloning reaching open-source SOTA (from ModelScope release info, not independently verified).

Why It Matters

  • Open-sourcing a complete model family spanning TTS, ASR, and audio understanding in one go under MIT, with Day-0 vLLM-Omni support, lowers the barrier for community deployment and further development.
  • The spec range from 0.5B to 126B-A6.5B covers both edge and high-capability scenarios, and combined with leaderboard results, forms a strong open-source alternative to closed voice models.

Episode 2 · Ant's Ming-3.0-tts tops open-weight TTS leaderboards with best pronunciation robustness (2026-10-20, 8 posts)

On October 20, Artificial Analysis released/updated several speech synthesis (TTS) leaderboards, and Ming-3.0-tts, open-sourced by Ant Group's Ling AGI team, emerged as the biggest winner in the open-weights camp. The model is a 0.5B-parameter dense model supporting preset voices and stateless voice cloning, with a 7.9B MoE variant whose weights are released under the MIT license.

Confirmed

  • In the Pronunciation Robustness evaluation, Ming-3.0-tts ranked first among all models with an overall score of 93.2%; the evaluation was conducted by human reviewers checking each model's reading of four categories of difficult text against pre-agreed standard pronunciations.
  • In the pronunciation robustness sub-categories, Ming-3.0-tts ranked first in "abbreviation expansion" with 96.5%, ahead of Eleven v4 (94.1%); Gemini led in the phrase category.
  • In the Controlled Voice Arena (where all models are compared using the same 8 cloned voices), Ming-3.0-tts ranked first among open-weights TTS with an Elo of 1069, ahead of Drama 3 Preview.
  • In the Provider Voice Arena (where each model uses its own native voices), Ming-3.0-tts scored an Elo of 1188, ranking 19th overall and third among open-source models; it performed best in the customer service (Elo 1230) and assistant (1216) scenarios.

Why it matters

  • With its lightweight 0.5B size and MIT open license, it surpasses commercial closed-source models on multiple leaderboards (e.g., the abbreviation expansion sub-score over Eleven v4), showing that open-source TTS is now competitive in pronunciation accuracy and voice controllability.
  • New evaluation dimensions such as pronunciation robustness, backed by human review, give the industry a more objective measure of technical capability than Elo-based subjective preference.