Ant Group's Ant Ling open-sources Ming-3.0, five MIT-licensed speech models: a 0.5B TTS that tops Artificial Analysis' open-weights voice-cloning arena, a 126B-A6.5B audio-understanding MoE and a dialect-aware ASR

Why it matters

【Key points】

On Oct 19 (15:54 UTC), Ant Group's Ant Ling team (inclusionAI) open-sourced Ming-3.0: five speech models in three families, MIT-licensed on Hugging Face and ModelScope.

Ming-3.0-tts comes as a 0.5B dense model and a "tiny" MoE built on Ling V3 (7.9B total, 1.3B active). It clones or designs voices from text, takes natural-language delivery instructions and 14 inline tags such as laughter and sighs, and covers seven languages plus Sichuanese and Cantonese; tiny also does multi-speaker dialogue.

Ming-3.0-audio-flash is a 126B-total, 6.5B-active MoE for recognition, long-recording summaries, spoken Q&A, sound and music understanding and tool requests, with a 9.3B (2.7B active) tiny version. Ming-3.0-asr-tiny (8.6B total, 2B active) handles Chinese, English and dialects, with hotword and context prompting. vLLM and SGLang had day-0 support.

On Oct 20, Artificial Analysis ranked Ming-3.0-tts the top open-weights model in its Controlled Voice TTS arena (Elo 1069, 13th overall) and first overall on Pronunciation Robustness (93.2%).

【Verifiable】

Hugging Face collection: https://huggingface.co/collections/inclusionAI/ming-30-6ad62ac04ca3a3f15e976f34

GitHub (TTS): https://github.com/inclusionAI/Ming-3.0-tts

GitHub (audio): https://github.com/inclusionAI/Ming-3.0-audio-flash

GitHub (ASR): https://github.com/inclusionAI/Ming-3.0-asr

Ant Ling on X: https://x.com/AntLingAGI/status/2112210811997540475

Artificial Analysis: https://x.com/ArtificialAnlys/status/2112389514404462613

【Our read】

1. A full MIT-licensed speech stack (synthesis, understanding, recognition) lets voiceover and transcription move from per-call APIs to self-hosted servers.

2. A 0.5B model topping an independent open-weights TTS ranking shows voice quality no longer needs scale; the 126B audio-flash is among the largest open audio models.

3. Apart from the Artificial Analysis ranks, benchmarks are Ant's own and no technical report is out yet.