Ant Group's Ming-3.0 Quietly Tops Open-Source Speech Benchmarks at 93.2%

Ant Group’s inclusionAI has released Ming-3.0-tts-0.5B, an MIT-licensed text-to-speech model with 500 million parameters. The weights arrived without a technical report, launch post, or formal announcement, but independent evaluations place the model at the top of two Artificial Analysis rankings for open-weight systems.

Artificial Analysis reports a 93.2% Pronunciation Robustness score, tying Ming with Fish Audio Drama 3 Preview for first place among 31 models. Eleven v4 scores 91.7%. The benchmark focuses on a recurring TTS failure mode: reading abbreviations, codes, unusual terms, and ordered character sequences correctly.

Three dates explain the rollout

Repository timestamps show a staggered release that separated the model weights from the code required to run them locally.

The weights therefore remained available for roughly four weeks before inclusionAI published its local inference tooling.

Two TTS checkpoints, different footprints

Ant released two checkpoints under the Ming-3.0 TTS name, and their labels describe different deployment profiles.

Stateless cloning conditions each generation on a supplied reference clip and creates no persistent speaker profile. Applications must provide or manage that reference audio for each request.

The TTS checkpoints form part of a five-model release spanning three Ming speech families:

• Ming-3.0-tts: 0.5B and Tiny 7.9B A1.3B models for speech synthesis
• Ming-3.0-audio: Flash 126B A6.5B for audio understanding and tool use, plus Tiny 9.3B A2.7B for efficient dialogue
• Ming-3.0-asr: Tiny 8.6B for automatic speech recognition

In the mixture-of-experts names, the value after “A” indicates the approximate number of parameters activated for each token. The release lists integrations for SGLang and vLLM, with weights distributed through Hugging Face and ModelScope.

A compact backbone drives 44.1 kHz audio

Ming-3.0-tts-0.5B combines a Qwen2.5-0.5B backbone, a diffusion-transformer head, and a separate audio tokenizer. It generates 44.1 kHz audio and supports seven languages plus Cantonese and Sichuanese.

Fourteen inline paralinguistic tags can trigger sounds such as laughter, sighs, and coughs within an utterance. Developers place a supported tag in the input text, and the model synthesizes the event in context alongside the spoken words.

The checkpoint occupies 2.96 GB in safetensors format. InclusionAI recommends an NVIDIA H800 and pins Python 3.10.15, PyTorch and torchaudio 2.7.1 with CUDA 11.8, and Transformers 4.46.3. The checkpoint size covers stored weights; inference also requires memory for activations, the audio stack, and runtime overhead. InclusionAI has not supplied consumer-GPU latency or minimum-VRAM figures in the release materials described here.

The benchmark split reveals its strengths

The Controlled Voice Arena tests every system with the same eight cloned voices. Ming leads the open-weight group there, ahead of Drama 3 Preview at 1,059 Elo and Breeze TTS 2 at 1,019.

The Provider Voice Arena evaluates each system with its own supplied voices. Ming trails Drama 3 Preview at 1,245 Elo and Breeze TTS 2 at 1,220 among open-weight entries. Its strongest reported results therefore come from pronunciation robustness and controlled voice cloning, with weaker placement for preset voice quality and some context-sensitive terms.

Arena Elo scores derive from comparative evaluations and can move as additional appearances accumulate. Teams should treat the rankings as a useful screening signal and test representative scripts, languages, voices, and failure cases before selecting a production model.

MIT licensing clears a commercial path

The MIT license permits commercial use, modification, redistribution, and self-hosted deployment while requiring preservation of the license notice. That flexibility distinguishes Ming from speech checkpoints limited to research or noncommercial use.

The model’s license and cloning support make several deployments practical:

• Self-hosted narration for audiobooks, podcasts, and video dubbing
• Private voice agents that keep audio within controlled infrastructure
• Game and animation pipelines generating large volumes of dialogue
• Multilingual interactive voice response systems using one checkpoint
• Offline or edge-oriented applications where network access is constrained

The MIT terms do not grant rights to a person’s voice, recordings, trademarks, or other input material. Production systems still need consent controls, provenance records, disclosure policies, and compliance with applicable biometric and synthetic-media laws. Dependencies and bundled assets may also carry separate licenses.

Self-hosting changes the cost model

Artificial Analysis lists SpeechifyAI’s hosted Simba-3.2 voice at $6.58 per million characters, the lowest price among the top ten systems in its speech arena. Ming replaces per-character API charges with GPU time, storage, engineering, monitoring, and capacity planning.

Break-even depends on utilization, batching, target latency, hardware prices, and operational staffing. A useful internal comparison should measure real-time factor, peak VRAM, concurrent request throughput, cold-start time, and total cost per generated hour on the hardware intended for production.

Production tests that matter

A deployment evaluation should cover the model’s published strengths and the operational details absent from the release materials:

1. Pronunciation: Test product names, abbreviations, dates, identifiers, URLs, multilingual passages, and domain-specific terms.
2. Voice consistency: Compare multiple reference clips, recording conditions, speaker ages, accents, and utterance lengths.
3. Performance: Record latency, real-time factor, VRAM use, batch scaling, and long-form stability on target GPUs.
4. Audio quality: Check pacing, prosody, clipping, background artifacts, and transitions around inline sound tags.
5. Language coverage: Evaluate every required language and dialect separately because aggregate rankings can hide uneven quality.
6. Serving compatibility: Validate the supplied scripts and SGLang or vLLM integrations against the production environment.
7. Safety controls: Add speaker consent, access restrictions, watermarking or disclosure where appropriate, and abuse monitoring.

Where Ming fits

InclusionAI has released MIT-licensed models across its Ling language, Ring reasoning, and Ming multimodal families. The Ming-3.0 speech rollout relied on model repositories and executable examples, with independent benchmarks supplying much of the public performance record.

Ming-3.0-tts-0.5B belongs on the evaluation list for teams that need self-hosted voice cloning, strong text normalization, permissive licensing, and a compact checkpoint. Products centered on polished preset voices or context-sensitive pronunciation should compare it directly with Drama 3 Preview, Breeze TTS 2, and hosted commercial systems using their own scripts and deployment hardware.