Ming-3.0-tts-tiny Has the Best Cloned Voice in the Family and Finishes Sixth in Korean

The best speaker-similarity score in Ant Group's new speech family belongs to the model with the fewest active parameters. Ming-3.0-tts-tiny is a 7.9B mixture-of-experts checkpoint that activates roughly 1.3B parameters per token, and on the Chinese Seed-TTS set it posts 0.781 speaker similarity — the highest figure anywhere in Ant's published tables, and higher than the 0.758 of its own smaller sibling. On Korean, the same model finishes sixth of the eight systems Ant benchmarked, at 6.000% character error rate, while that 0.5B sibling reaches 1.728%. Both numbers come from the same vendor harness, run by the same team, on the same day. The distance between them is the most useful thing in the release. 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.

None of this was announced. The checkpoint went up on Hugging Face on September 24, 2026, two days after the repository itself was created; the inference code, the example scripts and the task cookbook did not arrive until October 19, when the GitHub repository was opened and its first commit landed at 08:58 UTC. At the time of writing the Hugging Face repository had 20 downloads and 3 likes, and the GitHub repository had five stars. So treat what follows as a reading of the artefacts rather than a report on a launch: everything here is checkable in the repositories, and the parts that are not are marked as such.

The Korean result is not a rounding error

Ant's multilingual table compares the tiny model against seven other systems on six languages. Read the whole row rather than the headline.

• German — 0.325% character error rate, first of the eight, ahead of the next best at 0.612%.

• Spanish — 0.971%, first, with the runner-up at 1.184%.

• Japanese — 3.427%, tied for second.

• French — 3.470%, second, but with a speaker similarity of 0.733 that does not place in the top three; the table says so in its own limitations section.

• Korean — 6.000%, sixth of eight, and the only language where the model is beaten by a system it otherwise outscores everywhere.

• Chinese and English — the two languages the model is actually built around, and the two where its speaker similarity leads.

The Korean figure is roughly three and a half times the error rate of Ming-3.0-tts-0.5B on the same language. Ant does not explain why, and the model card does not speculate. What can be said without guessing is structural: this is a sparse model, and sparse models route tokens through a subset of experts, so a language that received less training signal in the mixture can degrade faster than it would in a dense model of comparable total size. That is a plausible mechanism, not a finding, and it should be treated as a hypothesis to test rather than an explanation. The practical reading is narrower and more defensible: if Korean is in your product, this is not the checkpoint to reach for until someone runs it on real Korean text.

Where 7.9 billion parameters actually go

The name says tiny; the file sizes say otherwise, and the gap is the architecture.

• Total parameters — the Hugging Face metadata reports 8,980,964,372 in BF16 plus 40,096 in F32, and the repository occupies about 18 GB. The release describes the model as 7.9B-A1.3B, meaning 7.9 billion total and about 1.3 billion active per token. That description was corrected in the model card on October 16, replacing an earlier wording, so quote the newer one.

• Backbone — Ling V3, the same family Ant uses for its larger text models, run as a mixture of experts rather than as a dense stack.

• Head — a diffusion transformer that turns the backbone's hidden states into audio latents, followed by a separate audio tokenizer that decodes them to waveform.

• Bundle — the checkpoint ships the backbone, the DiT head, the audio tokenizer, a CAM++ speaker encoder and the text tokenizer in one folder. You are downloading a pipeline, not a language model with an adapter bolted on.

• Output and voices — the same broad control surface as the rest of the family: zero-shot cloning from a short reference clip, voice design from a natural-language description, free-form style control, 100-plus built-in Chinese voices and five English ones, and 14 inline paralinguistic tags for laughter, sighs, breath and hesitation, plus explicit emotion, speed, pitch and volume settings.

• Languages — Chinese, English, German, Spanish, French, Korean and Japanese, with Sichuanese and Cantonese available when given a reference clip in the dialect.

• Licence — MIT for weights and code, with a usage note asking you not to impersonate people without consent and to label synthetic audio. That note is a request, not a licence condition.

The odd consequence is that "tiny" is a statement about compute per token, not about disk or VRAM. You spend the memory of a 7.9B model and get the speed of a 1.3B one, which is a reasonable trade for batch synthesis and a poor one for a laptop. Ant's own environment pins an H800 as the recommended card, and nothing in the release suggests otherwise.

The five columns it wins

Every figure in this section is Ant's own, measured on Ant's harness with Ant's recommended inference settings, and the model card is the only place it exists. The labels matter more than usual here, because these are the numbers that will be quoted elsewhere without them.

• Speaker similarity, Chinese zero-shot — 0.781, the best in the comparison set, with a word error rate of 0.820%. English reaches 0.728 similarity at 1.666% error rate, also the best similarity in its column.

• Paralinguistic single-event, NVBench — the model leads all six single-tag categories: Chinese 2.80 error / 17.23 similarity / 3.20 score, English 5.86 / 32.95 / 7.05. On multi-event sentences it leads the English column at 5.12 / 46.13 / 7.68 and sits mid-pack in Chinese at 4.49 / 32.88 / 5.81.

• Dialect — WSC-Eval-TTS easy set 2.53% character error, 0.721 similarity, 86.09% accuracy; hard set 3.11%, 0.738, 88.94%. Both are firsts in Ant's table. Cantonese reaches 7.40% / 0.605 / 92.90 on the base set and 8.67% / 0.653 / 95.16 on the coverage set.

• Emotion — 74.00% average on the Chinese CV3 set and 30.67% on CV3N, first on both of the Sad subsets and unremarkable elsewhere.

• Voice design — 75.37% average on InstructTTSEval-ZH, second overall, with attribute-prompt scores of 85.10, description 79.90 and role-play 61.10. Role-play is the weak one: fourth place, and the model card says so.

The basic-attribute table is the least glamorous and probably the most useful: 93% on age, 93% on gender, 78.70% on pitch, an average of 88.23% with 0.274% word error rate and 0.742 similarity. Pitch is the weakest of the four controls, which matches what every other system in this class reports.

The podcast number is the real reason to notice this

The 0.5B model cannot do multi-speaker podcast generation. The tiny one can, and on Ant's podcast table it is the best content-accuracy result in the comparison.

• Ming-3.0-tts-tiny — 2.03% character error rate, 0.561 conditional speaker similarity, 2.12 UTMOS.

• SoulX-Podcast — 2.20% / 0.599 / 2.09.

• ZipVoice-Dia — 3.39% / 0.553 / 2.24.

• FireRedTTS2 — 3.34% / 0.512 / 1.90.

• MOSS-TTSD — 8.62% / 0.421 / 1.70.

• VibeVoice-1.5B — 12.87% / 0.455 / 1.74.

• MoonCast — 27.43% / 0.441 / 1.76.

First on error rate by a small margin, second on both similarity and naturalness, and the naturalness gap is the interesting one: ZipVoice-Dia is 0.12 UTMOS ahead while being 1.36 points worse on accuracy. That is a real trade rather than a rounding difference, and it suggests the tiny model's dialogue prosody is the next thing to improve rather than its transcription. As with everything else in this section, no one outside Ant has run a podcast through it and published the result.

What the repositories do not establish

Four gaps are worth naming before anyone puts this in a product.

There is no independent score for this checkpoint. Artificial Analysis does list a Ming-3.0-tts entry on its speech arena — 1,188 Elo on the Provider Voice board and 1,069 on Controlled Voice as of October 19 — but that entry is linked to the 0.5B model, not to this one. The GitHub README reproduces those arena figures in its results section without that distinction, which is the kind of detail that turns into a wrong claim two reposts later. Ming-3.0-tts-tiny has no separate arena result, no measured throughput, and no third-party word error rate.

There is no technical report. The roadmap lists it as an unchecked box, and on October 16 the model card was edited specifically to remove a reference to one. That leaves the architecture at the level of a diagram and a paragraph: enough to run the model, not enough to reproduce the training claims.

There is no hosted endpoint and no price. The only way to call this is to allocate a GPU and run the code, which is a different proposition from comparing it against a per-character rate.

And there is no statement of intent. Twenty downloads and three likes are not a signal of quality — they are a signal that nobody has been told. No announcement means no support commitment, no deprecation policy, and no indication of whether a hosted service follows.

Running it, if you want to

The Hugging Face repository is a checkpoint drop, and its card says plainly that inference requires the GitHub project. That is the reason the October commit matters more than the September upload. The environment is pinned to Python 3.10.15, PyTorch and torchaudio 2.7.1 with CUDA 11.8, and transformers 4.46.3, with an H800 recommended; the entry point is infer_tiny.py against the ming_tiny.MingTinyTTS class, with a prompt protocol named thinking_v1. The released defaults are a classifier-free guidance scale of 2.0, zero temperature, ten flow steps, a sigma of 0.25 and a seed of 1234 — 1235 for the podcast task — and the repository ships a separate script per task rather than one flag-driven entry point. vLLM-Omni support is on the roadmap and not yet checked, so there is no high-throughput serving path today.

If you are evaluating this for a product, the sensible staging is the one you would use for any unproven open-weights checkpoint: keep it off the critical path, run it against your own audio in the languages you actually ship, and only then decide. OrcaRouter does not route Ming-3.0-tts-tiny, and nothing here should be read as claiming otherwise — the catalogue carries no Ming speech model. What it does carry is the rest of a voice system: openai/tts-1, openai/tts-1-hd and openai/gpt-4o-mini-tts plus the Ge​mini TT​S previews at provider list price with no markup added, and more than 200 text models behind the same key for the half of the pipeline that writes the script or drives the agent. Automatic failover across providers is what keeps the hostable text half up while you decide whether the self-hosted voice half has earned its slot.

The next real signal will not be another vendor table. It will be someone outside Ant running this on Korean, or on a podcast, and publishing what broke. Until that exists, the defensible summary is narrow: a sparse, MIT-licensed, nine-language voice model with the best cloned-voice score in its own family, a genuine multi-speaker podcast result, a Korean column that needs explaining, and an announcement that has not happened.