Ming-3.0-tts Tops the Pronunciation Arena, and the Rest of Its Debut Is More Interesting

At 93.2%, Ming-3.0-tts and Fish Audio Drama 3 Preview are separated by one hundredth of a percentage point — 1,677 correct spans out of 1,800 for Ant Group's model against 1,785 out of 1,916 for Fish Audio's — which is a tie in everything but the decimal. Both sit above Eleven v4 at 91.7%, which is the number most people will quote. The tie is genuine and worth reporting straight, but it is the least informative thing on the board. What makes Ming-3.0-tts worth a piece this week is not that it reached first place on Artificial Analysis's English Pronunciation Robustness leaderboard. It is that the board finally exists for this model at all, and that the four sub-scores underneath the headline say something the headline cannot. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Those sub-scores are lopsided. Ming-3.0-tts leads all 31 measured systems on Preserving Exact Sequences, at 90.6%, and on Expanding Shorthand, at 96.5%. It also ranks 21st of 31 on Standalone Terms, at 88.5%, which is the category most listeners would call "mispronouncing names." One aggregate score hides a model that reads .env.local and "Chapter XVII" better than anything else on the board and stumbles on ordinary proper nouns more often than a G​emini or an ElevenLabs system does. Whether that is a win depends entirely on what your text is made of.

What actually changed, and what did not

Ant Group did not launch this model this week, and no article should pretend otherwise. The timeline in the repositories is public and slightly awkward:

• The Hugging Face repositories for Ming-3.0-tts-0.5B and Ming-3.0-tts-tiny were created on 22 September 2026, with a placeholder card.

• The 0.5B checkpoint — 2.96 GB of safetensors, MIT-licensed — went up on 24 September.

• The GitHub repository with the inference code, the task cookbook and the example scripts appeared on 19 October, which is also the date Ant's README gives as the release month and the date Artificial Analysis stamps on its model row.

So the weights are four weeks old and the usable release is one day old. What changed on 19 October is not that a model shipped. It is that the inference code shipped, and that the independent measurements arrived alongside it. Before that date there was a checkpoint with no published way to run it and no third-party number attached; anyone writing about it was quoting Ant's own tables.

That distinction matters for how you read the 93.2%. Artificial Analysis runs its own harness and its own human reviewers; these are not Ant's evaluations reproduced. The evaluation dates and sample counts are listed on the board, the confidence intervals are published, and the reviewers' attention checks are described in the methodology. This is the first time anyone outside Ant has put a number on this model, and it happens to be a good one.

Three boards, three different answers

Artificial Analysis runs more than one text-to-speech board, and Ming-3.0-tts does not dominate all of them. The differences are large enough that quoting whichever one flatters the model is a form of misrepresentation.

• Controlled Voice, where every system clones the same eight reference voices (four US, four UK) and only the synthesis is compared — Ming-3.0-tts scores 1,068 Elo, first among open-weights models and 12th overall. Fish Audio Drama 3 Preview is second among open weights at 1,059.

• Provider Voice, where each system uses its own native voices — Ming-3.0-tts scores 1,188 Elo, 18th overall and third among open-weights models, behind Fish Audio Drama 3 Preview at 1,247 and Breeze T​TS 2 at 1,221. Eleven v4 Turbo leads the whole board at 1,330.

• Pronunciation Robustness, English, human-judged — Ming-3.0-tts at 93.2%, first overall by a hair over Fish Audio Drama 3 Preview at 93.2% and ahead of Eleven v4 at 91.7%.

The pattern is consistent: as soon as the voices stop being held constant, Ming-3.0-tts loses ground. That is the signature of a model whose cloning is stronger than its stock voice set — which is exactly what a checkpoint with no hosted voice catalogue should look like. Ant ships 100-plus Chinese built-in voices and five English ones; three of those English voices are what the Provider Voice board actually heard, against ElevenLabs' and Cartesia's carefully produced catalogues. On cloned voices the playing field levels and the model moves up five places.

Note also that the numbers have already drifted from Ant's own README. The repository table says 1,069 Controlled and 13th overall, and 1,188 Provider and 17th overall. The board today reads 1,068 and 12th, and 1,187.8 and 18th. Votes keep arriving; the line moves. Anyone citing these figures should date them.

The 93.2 is two wins and two losses

Pronunciation Robustness is not one test. Artificial Analysis splits it into four categories, each with its own denominator, and the split is where this model becomes legible.

• Preserving Exact Sequences — codes, file paths, email addresses, identifiers. Ming-3.0-tts scores 90.6%, first of 31, with the next best being Fish Audio Drama 3 Preview at 86.3% and Eleven v4 at 78.8%. This is the category that decides whether a support bot reads an order number correctly, and the margin is not close.

• Expanding Shorthand — numbers, dates, units and notation. Ming-3.0-tts scores 96.5%, first of 31, ahead of Eleven v4 at 94.1% and Fish Audio Drama 3 Preview at 93.6%.

• Contextually Appropriate — heteronyms such as "a wound that is bandaged" against "a bandage that is wound." Ming-3.0-tts scores 91.0%, 13th of 31. Gemini 3.8 Flash T​TS leads at 98.0%.

• Standalone Terms — brand, place and technical names. Ming-3.0-tts scores 88.5%, 21st of 31. Fish Audio Drama 3 Preview leads at 95.9%.

An aggregate that ties for first while ranking 21st in one of its four components is not a paradox; the two categories it wins carry the largest denominators (1,084 of the 1,800 judged spans are shorthand and sequence cases). Ant's own README leans on precisely these two wins, calling out Expanding Shorthand at 96.5% and Preserving Exact Sequences at 90.6%, and the independent run confirms both. It just does not mention the 21st place, which is the row a voice-agent team should look at before shipping a model that has to say customer surnames.

The 21st place is also the most plausible explanation for a detail in Ant's model card: the English voice set is five voices against 100-plus in Chinese. English proper-noun coverage is a data problem before it is a modelling problem, and the term-pronunciation number is what that gap looks like when someone measures it.

The half of the family that is not on the board

Ming-3.0-tts-tiny, the 7.9B-total, 1.3B-active mixture-of-experts variant built on the Ling V3 backbone, has no row on any of these leaderboards. The arena entry labelled Ming-3.0-tts links to the 0.5B checkpoint, and every number above belongs to that checkpoint. Tiny's published results are Ant's own — Chinese word error 0.820% with speaker similarity 0.781, German 0.325%, Spanish 0.971%, Korean 6.000% — and nobody independent has reproduced them. Treat them as vendor claims; the embarrassing Korean figure is at least evidence that Ant is publishing the table rather than a highlight reel.

There is one place where tiny looks genuinely ahead of its sibling on Ant's harness: multi-speaker podcast generation, which the 0.5B model does not do at all. That is a product capability rather than a benchmark edge, and it is not on the arena.

Running it, and what that costs

The 0.5B checkpoint is a 2.96 GB download — backbone, DiT patch head, audio tokenizer, speaker encoder and text tokenizer — under MIT, outputting 44.1 kHz audio. Ant tested it on an H800-80GB and an H20-96G with PyTorch 2.6, CUDA 12.4 and transformers 4.46.3, and the repository includes a Dockerfile. Inference goes through infer.py or the MingTTS Python class; the Hugging Face repository is checkpoint distribution only and will not run on its own.

Artificial Analysis measured the hosted Ant endpoint at 44.3 characters per second of generation time, with a median generation time of 2.56 seconds. For scale: a million characters of finished narration is roughly 6.3 hours of pure generation on that configuration, before model loading or queueing. There is no listed price because there is no hosted commercial endpoint; you are buying GPU hours, not characters.

This is where a routing layer is genuinely useful without pretending to host the model. OrcaRouter does not carry Ming-3.0-tts — the catalogue has no Ming speech model at all — and the model is reached by cloning the repository and running it yourself. What we do carry is the text half of the same pipeline: more than 200 models behind one key at provider list price with no markup added, automatic failover between providers, and a routing DSL for composing them. If you are building a narration or agent pipeline, the model that writes the script is the part you can consolidate today; the voice model is the part you self-host.

What to watch from here

Three things would change this picture, and none of them has happened yet.

Ant's roadmap still lists the technical report as unchecked. Until it lands, the architecture description is a paragraph and a diagram, and the evaluation tables are the only detailed evidence about how the model behaves. A report would also explain the term-pronunciation gap, which is currently a hypothesis.

Provider Voice has heard three native voices. That is a thin sample for a model that ships a hundred in Chinese and five in English, and the 18th place will move as more voices are scored — in which direction is not knowable from here.

And the tiny checkpoint remains unmeasured by anyone independent. If it does get a row, the interesting question is whether a Ling V3 mixture-of-experts at 1.3B active per token keeps the sequence-fidelity advantage that makes the 0.5B model distinctive, or trades it for the speaker similarity that Ant's own tables give it. On the vendor numbers, tiny is the better cloned voice. On the only independent numbers that exist, the 0.5B is the one that has been tested, and it passed the part of the test that most people do not think to run.