Ant Group's Ming 3.0 Tops 31 Open Speech Models at Just 0.5B
Ant Group's Ling team open sourced five Ming 3.0 speech models under MIT, with the 0.5B TTS topping open-weight voice benchmarks.
Takeaways
- Ant Group open sourced five Ming-3.0 speech models under MIT license across TTS, audio understanding, and ASR.
- Ming-3.0-tts-0.5B tops the Artificial Analysis Controlled Voice open weights leaderboard with 1,069 Elo.
- Pronunciation Robustness score of 93.2% ties Fish Audio Drama 3 and beats ElevenLabs v4.
- Supports 7 languages, Sichuanese and Cantonese dialects, 14 paralinguistic tags, voice cloning, and natural-language voice design.
- Audio Flash is 126B MoE with 6.5B active, handling long recordings, music, ambient sound, and tool calls.
- Available on Hugging Face and GitHub with day-0 SGLang and vLLM support.
Ant Group’s Ling team released the Ming 3.0 speech family on October 20, publishing five open-weight models for speech synthesis, audio understanding, conversation, and transcription. The repositories are available through Hugging Face and ModelScope under the MIT license, which permits commercial use, modification, and redistribution.
Developers can deploy each model separately or combine them into a voice pipeline that accepts microphone input, interprets speech, invokes tools, and generates a spoken response. The 0.5B checkpoint leads the release with competitive text-to-speech results from a relatively small dense model.
Five checkpoints, three workloads
The family assigns synthesis, audio reasoning, and transcription to specialized checkpoints. Two models use dense architectures, which activate every parameter during inference. The larger mixture-of-experts models activate only a subset of their total parameters for each input, reducing compute relative to their headline size.
- Ming-3.0-tts-0.5B: Dense text-to-speech model for voice cloning, voice design, multilingual synthesis, and paralinguistic cues such as laughter and sighs.
- Ming-3.0-tts-tiny: A 7.9B-parameter mixture-of-experts model with 1.3B active parameters, designed for long-form narration, stories, and voice assistants.
- Ming-3.0-audio-flash: A 126B-parameter model with 6.5B active parameters for long recordings, spoken questions, music, ambient audio, and tool calls.
- Ming-3.0-audio-tiny: A 9.3B-parameter model with 2.7B active parameters for lower-cost spoken dialogue.
- Ming-3.0-asr-tiny: An 8.6B speech-recognition model covering seven languages, dialects, hotwords, and contextual transcription.
A small TTS model with granular controls
Ming-3.0-tts-0.5B combines a Qwen2.5-0.5B language-model backbone, a diffusion transformer acoustic head, an audio tokenizer, a speaker encoder, and a text tokenizer in one checkpoint. It generates 44.1 kHz audio in Chinese, English, German, Spanish, French, Korean, and Japanese, with additional controls for Sichuanese and Cantonese.