Ming-3.0-tts vs Ming-3.0-tts-tiny: The Smaller Checkpoint Is the Better Download, and the Bigger One Is the Only One Anyone Has Measured

Ant Group shipped two speech models under one name, and the naming is the first thing that will mislead you. Ming-3.0-tts-the-default is the 0.5B dense model: 2.96 GB on disk, a Qwen2.5-0.5B backbone, fast, cheap, and the only one of the pair with an independent score. Ming-3.0-tts-tiny is the larger of the two by an order of magnitude — a 7.9B-parameter mixture of experts with 1.3B active per token, built on the Ling V3 backbone, 18 GB of sharded weights — and it is called tiny because of what it costs to run, not what it costs to download. Pick by the label and you will get the wrong file. 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.

The second thing that will mislead you is the leaderboard. Artificial Analysis carries a Ming-3.0-tts row at 1,188 Provider Voice Elo and 93.2% pronunciation robustness, and that row links to the 0.5B checkpoint. None of it belongs to tiny. Tiny's numbers are Ant's own tables, published in the model card and the GitHub README, and no third party has reproduced any of them. So the honest comparison has two halves: one model with independent evidence and no podcast mode, one model with vendor evidence and a capability the other lacks. Which half you need is a question about your workload, not about which model is better.

Same architecture family, two very different machines

Both models share the generation framework Ant describes in the README: text and task-specific control tokens condition a language model, which drives a diffusion-transformer patch head that emits audio tokens. The difference is the language model underneath. The 0.5B uses a dense Qwen2.5-0.5B; tiny uses a sparse Ling V3 mixture of experts at 7.9B total with roughly 1.3B activated per token.

• Download size — Ming-3.0-tts 2.96 GB vs Ming-3.0-tts-tiny about 18 GB across 30 shards

• Parameters — 0.5B dense vs 7.9B total / 1.3B active MoE

• Backbone — Qwen2.5-0.5B vs Ling V3

• Runtime — PyTorch 2.6 + CUDA 12.4 on Python 3.8-era stack vs Python 3.10.15 + PyTorch/torchaudio 2.7.1+cu118

• Hardware — tested on H800-80GB and H20-96G vs H800 recommended, SGLang-Omni can serve on one GPU

• Multi-speaker podcast generation — not supported vs supported

• Licence — MIT vs MIT

• Independent arena score — yes, 1,188 Provider / 1,068 Controlled / 93.2% pronunciation vs none

The two checkpoints do not share a Python environment. That is not a documentation quirk; it means a team that wants to evaluate both on the same box is maintaining two virtualenvs, and a team that wants to A/B them in production is maintaining two deployment images. If your plan is "start with the small one and upgrade later," budget for the migration as a real piece of work, not a config change.

Where tiny is genuinely better, on Ant's own numbers

Speaker similarity is the clearest and most consistent win, and it is the number that decides whether a cloned voice sounds like the person.

• Seed-TTS Chinese speaker similarity — Ming-3.0-tts 0.758 vs tiny 0.781

• Seed-TTS English speaker similarity — 0.691 vs 0.728

• German similarity — 0.753 vs 0.791

• Spanish similarity — 0.763 vs 0.797

• French similarity — 0.675 vs 0.733

• Japanese similarity — 0.765 vs 0.780

• Korean similarity — 0.793 vs 0.797

Tiny wins every language on similarity, usually by three to six points, and the pattern holds on the control benchmarks too: 74.00% average on CV3 emotion against 72.33%, and 75.37% on voice-design accuracy against 72.29%. Those are the two things a customer notices without being told — does it sound like me, and did it do what I described. On both, the 7.9B MoE is ahead.

Content accuracy splits the other way, and the split is not symmetric.

• Seed-TTS Chinese word error — 0.800% vs 0.820%. Effectively a tie; the dense model is marginally ahead.

• Seed-TTS English word error — 1.584% vs 1.666%. Dense ahead.

• German — 0.365% vs 0.325%. Tiny ahead.

• Spanish — 1.099% vs 0.971%. Tiny ahead.

• French — 4.264% vs 3.470%. Tiny ahead by a wide margin.

• Japanese character error — 3.765% vs 3.427%. Tiny ahead.

• Korean character error — 1.728% vs 6.000%. Dense ahead by a factor of three and a half.

Korean is the outlier that should stop anyone from reading "tiny is better at multilingual" off the table. Ant's own card flags it: Korean has the largest multilingual accuracy gap in the tiny model, and the 6.000% figure is more than triple what the half-billion model manages. A team building for the Korean market has a clear answer and it is not the bigger model.

The capability that is not a benchmark

Ming-3.0-tts-tiny generates multi-speaker podcast dialogue from a script with speaker references. Ming-3.0-tts does not. That is a product capability, not a score, and it does not appear on any leaderboard — which is exactly why it is easy to miss when comparing spec sheets that both list "voice cloning" and "voice design."

Ant's podcast evaluation, on its own harness, puts tiny at 2.03% character error rate with 0.561 speaker similarity and 2.12 UTMOS against ZipVoice-Dia at 2.24 UTMOS, FireRedTTS2 at 3.34% CER and MOSS-TTSD at 8.62%. The UTMOS gap is the one to note: tiny leads on content accuracy and trails the best comparison on naturalness, which is the same trade-off the speaker-similarity numbers hint at from the other direction. Podcast generation rewards turn-taking and speaker consistency more than it rewards a single perfect clip, and that is the shape of tiny's advantage.

If you need two people talking, there is no comparison here — one model does it and one does not. Everything else in this article is about the case where both can do the job.

What the independence gap actually costs you

The 0.5B model's arena position is worth understanding before you decide how much weight to put on it. It scores 1,188 Provider Voice Elo, 18th of 97, and 1,068 Controlled Voice, first among open-weights models and 12th overall. On pronunciation it is at 93.2%, tied in practice with Fish Audio Drama 3 Preview and ahead of Eleven v4. Those are independent measurements with published sample counts and confidence intervals.

What that buys you is calibration. You know roughly where the dense model sits against ElevenLabs and Cartesia, because the same lab measured all three on the same board. For tiny, you know only what Ant says, measured on Ant's harness with Ant's settings, and Ant is the party with an interest in the result. This is not an accusation — the tables are unusually candid, including the Korean number and a list of limitations that names pitch control and role-play voice design as weak spots. It is a statement about what a number can support. A vendor table tells you the model is competitive on the axes the vendor chose to measure; an arena row tells you where it lands against systems you already use.

There is a specific thing the independent board does not tell you either: it measured one voice. Artificial Analysis notes that each model is evaluated on the pronunciation board with a single voice, and Ming's row uses Ant's endpoint with three native voices available. Whether the 93.2% holds across the cloned voices a product would actually ship is unknown, and it is the first thing worth testing yourself.

Choosing, in practice

The decision tree is short.

Take Ming-3.0-tts (0.5B) if you want the smallest viable download, if you are deploying for Korean, if your hardware is a single 80 GB card you also use for other things, or if you need a model whose behaviour has been measured by someone other than its author. At 2.96 GB and 44.3 characters per second on the independent harness, it is a serviceable narrator that outruns real-time speech on one GPU.

Take Ming-3.0-tts-tiny if speaker similarity is the product — a cloned brand voice, a personalised read-aloud, anything where the listener's first judgement is "does that sound like them" — or if you need podcast dialogue, or if French, German, Spanish or Japanese accuracy is the constraint. Budget 18 GB of weights, a separate Python 3.10 environment and an H800-class card.

Take neither, yet if you need an independent number before committing production capacity to tiny. The arena row does not exist, the technical report is still unchecked on the roadmap, and vLLM-Omni acceleration is marked done in the README with no published recipe. Those three gaps are the same three things that would make tiny easy to recommend.

One operational note that applies to both: neither checkpoint streams. There is no published time-to-first-audio for either model, and the arena's throughput figure is generation over a completed request. If your application is conversational, this family is not the tool regardless of which size you pick.

The routing question, which is different here

Most comparisons on this site end with a paragraph about consolidating access. This one cannot, because OrcaRouter does not host either Ming checkpoint — the catalogue carries no Ming speech model, and both are reached by cloning Ant's GitHub repository and running the weights yourself. Saying otherwise would be a lie a reader could check in thirty seconds.

What is true is narrower and still useful. A voice pipeline is usually text first: something writes the script, something translates it, something checks it against a house style. That text stage can sit behind one key — more than 200 models at provider list price with no markup added, automatic failover when a provider degrades, and a routing DSL for sending a task to a different model based on what the request contains. The speech stage in this comparison stays on your hardware either way, and the self-hosted path is what makes the choice between the two checkpoints reversible: MIT weights do not expire, and a model you already have on disk is one you can keep serving while you evaluate its replacement.

What would settle it

Three developments, in order of how much they would move the recommendation. A tiny row on the arena would tell you whether the similarity advantage survives blind listener comparison, which is the only test that has not been run. The technical report would explain why a 1.3B-active MoE gains speaker similarity and loses Korean accuracy, and whether that is a data split or an architectural consequence. And a published vLLM-Omni recipe would make the 18 GB checkpoint practical on rented hardware rather than a research deployment.

Until then, the useful framing is that Ant shipped two models for two jobs and gave them names that make the smaller job sound like the smaller file. It is not. The 0.5B is the one you can measure, deploy and defend in a design review; the 7.9B is the one that sounds more like the person you cloned, in six of seven languages, if you can afford to take that on trust.