Ming-3.0-tts-tiny vs Qwen-Audio-3.0-TTS: One Has a Rate Card, the Other Has a Repository

The comparison everyone wants is two scores side by side, and it cannot be made honestly today. Ming-3.0-tts-tiny is a 7.9-billion-parameter mixture-of-experts speech checkpoint from Ant Group's inclusionAI, released as weights under MIT with no hosted endpoint and no independent evaluation. Qwen-Audio-3.0-TTS is a hosted synthesis endpoint on A​libaba Cloud Model Studio, launched on July 20, 2026, sold in a Flash tier and a Plus tier, with no weights anywhere and no published error-rate table. One side has a leaderboard position and a price. The other has a downloadable pipeline and a vendor benchmark suite. The useful work here is not picking a winner from those two artefacts; it is working out which numbers may be placed next to each other, and which may not. 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 dates matter, because neither product arrived the way a launch post would suggest. Ming-3.0-tts-tiny's Hugging Face repository was created on September 22, 2026 and the checkpoint itself was uploaded on September 24. For three and a half weeks it was a folder of weights with no way to run them: the inference code, the one-click examples and the task cookbook only landed on GitHub on October 19 at 08:58 UTC, in nine commits across a single afternoon. Ant has not announced it, there is no blog post, no model card video and no technical report. Qwen-Audio-3.0-TTS, by contrast, had a real launch and is still on A​libaba's rate card — and the tier that matters for quality, Plus, is still the top of A​libaba's synthesis ladder even after the 3.1 generation arrived in September.

The leaderboard row is the wrong checkpoint, and the repository does not say so

Start here, because this is the mistake the first wave of comparison posts will make. Artificial Analysis maintains two speech boards, and both carry a model called Ming-3.0-tts. As of October 19 it sits at 1,188 Elo on the Provider Voice board and 1,069 on the Controlled Voice board.

That entry is the 0.5B model, not this one. Ming-3.0-tts-tiny has no arena row at all — not ranked, not unranked, not a preview entry. The GitHub README reproduces the 1,188 and 1,069 figures in its results section without distinguishing which checkpoint produced them, which is exactly how a number acquires the wrong label two reposts later.

So when you see "Ming-3.0-tts 1,188 vs Qwen-Audio-3.0-TTS-Plus 1,266", what is actually being compared is Ant's half-billion-parameter dense model against A​libaba's largest hosted tier, with the model in the title of this article absent from the board. The 78-point gap is real for the 0.5B. It says nothing about the tiny MoE, in either direction. The tiny model might be better; nobody has put it in front of a preference panel.

Absence is not a low score. It is an unscored model, and the distinction is the whole reason this section exists.

What each side actually is

Strip the marketing and the two objects differ in kind before they differ in quality.

• Form — Ming-3.0-tts-tiny is a checkpoint you download: roughly 18 GB, 8,980,964,372 parameters in BF16 plus 40,096 in F32, containing the Ling V3 backbone, a diffusion-transformer head, an audio tokenizer, a CAM++ speaker encoder and the text tokenizer. Qwen-Audio-3.0-TTS is an HTTP endpoint that returns audio; there is nothing to download and no licence to audit.

• Parameters — 7.9B total with about 1.3B active per token, a sparse mixture of experts, against a hosted model for which A​libaba publishes no parameter count.

• Licence — MIT on both weights and code, with a usage note asking you not to impersonate people without consent and to label synthetic audio. On the hosted side, commercial API terms.

• Languages — Chinese, English, German, Spanish, French, Korean and Japanese, plus Sichuanese and Cantonese when given a dialect reference clip. Qwen-Audio-3.0-TTS lists 16 languages and 20 Chinese dialect regions.

• Control surface — zero-shot cloning, voice design from a description, free-form style instructions, 14 inline paralinguistic tags, explicit emotion, speed, pitch and volume settings, and a multi-speaker podcast task. Qwen-Audio-3.0-TTS exposes cloning plus roughly 86 inline delivery tags and instruction-style control.

• Price — no listed rate on the Ming-3.0-tts-tiny side; you pay for GPU time. Qwen-Audio-3.0-TTS-Plus is ¥1.4 per 10,000 characters on A​libaba's Beijing list, which Artificial Analysis normalises to $19.31 per million characters.

• Independent score — none for Ming-3.0-tts-tiny. 1,266 Elo on Provider Voice and 1,126 on Controlled Voice for Qwen-Audio-3.0-TTS-Plus, both with confidence intervals that do not touch the Ming-3.0-tts row's 1,188 and 1,069.

Two kinds of evidence that do not convert into each other

The arena measures preference. Ant's tables measure error and similarity. Neither transforms into the other, and the temptation to treat an Elo gap as a quality multiplier should be resisted.

What Qwen-Audio-3.0-TTS gives you is a crowd's relative judgement of its native voices, with sample counts and intervals behind it. What Ant gives you is a laboratory's measurement of its own model against a fixed test set, with the harness, the reference clips and the scoring code all chosen by the same team that built the model. Both are legitimate. Neither is a substitute for the other, and only one of them exists for the model in this article's title.

Ant's suite is broad, and it is worth knowing what is in it before you dismiss it. On the Chinese Seed-TTS set the tiny model posts 0.820% word error at 0.781 speaker similarity — the best similarity figure anywhere in Ant's published tables, including its own 0.5B sibling's 0.758. English reaches 1.666% at 0.728. German lands at 0.325% character error and Spanish at 0.971%, both first in Ant's eight-system comparison. Japanese is 3.427%, French 3.470%. On NVBench's single-event paralinguistic test it leads all six categories, at 17.23 PCER in Chinese and 32.95 in English. The dialect suites come in at 86.09% accuracy on WSC-Eval-TTS easy and 92.90 on WSYue base, the emotion set at 74.00% average on CV3, and voice design at 75.37% on InstructTTSEval-ZH. All of it is Ant's own harness, run by Ant, reported by Ant, and reproduced by nobody.

The honest summary is that Qwen-Audio-3.0-TTS has been judged by strangers and Ant has been judged by its author. That is not a knock on either; it is a statement about what the two numbers are.

The one head-to-head Ant does publish, and it is against itself

Amid that suite there is a result that behaves like a real comparison, because the competitor is Ant's own 0.5B model and the difference is large enough to survive a different harness.

On Korean, Ming-3.0-tts-tiny posts a 6.000% character error rate. The 0.5B posts 1.728%. Same team, same test set, same reporting day — a factor of roughly three and a half, in the direction nobody expected from the larger model. The model card does not explain it, and the limitations section lists Korean among the known weak points without offering a mechanism.

A structural explanation is available and should be labelled as a hypothesis rather than a finding: sparse models route each token through a subset of experts, so a language that received less training signal inside the mixture can degrade faster than it would in a dense model of comparable total size. That is plausible. It is not measured. What is defensible is the operational conclusion, which is narrower: if Korean is in your product, this checkpoint is not the one to reach for until someone runs it on real Korean text and publishes the result.

Qwen-Audio-3.0-TTS covers Korean within its 16 languages, and A​libaba publishes no per-language error rate to hold against Ant's 6.000%. So even the weak spot cannot be converted into a comparison. It can only be read as a warning.

A million characters, priced twice

This is where the two products stop being abstractions.

Qwen-Audio-3.0-TTS bills per character. At the arena's normalised $19.31 per million characters, a million characters of finished narration costs $19.31, and a hundred million costs $1,931. The number is linear, predictable, and available before you write any code. Throughput on the same independent harness is 25.61 characters per second of generation time, which is a concurrency question rather than a billing one at the API's scale.

Ming-3.0-tts-tiny bills in GPU-hours, and the arithmetic has a hole in the middle of it. The 0.5B sibling was measured at 42.1 characters per second; the tiny model was not measured at all. You cannot borrow its sibling's throughput, because the tiny model activates roughly two and a half times as many parameters per token and occupies about 18 GB rather than 2.96 GB. Ant's recommended card is an H800, and the pinned environment — Python 3.10.15, torch and torchaudio 2.7.1 with CUDA 11.8, transformers 4.46.3 — is the only performance claim in the release.

So do the calculation in the open instead of pretending it is settled. At 20 characters per second, a million characters is about 14 hours of generation; at 10 characters per second it is about 28. Multiply those hours by your own GPU-hour cost, add model load time, queueing and retries, and compare with $19.31. Below a few million characters a month, the hosted rate usually wins on total cost once someone's time is in the figure. Above it, on hardware you already own and keep busy, the checkpoint's marginal cost per hour is the same as its first — which is the structural advantage no rate card can match.

One thing the price comparison should not hide: the two models are not the same product at any price. Qwen-Audio-3.0-TTS's Plus tier is a finished voice catalogue with 16 languages; Ming-3.0-tts-tiny is a pipeline you operate, whose output quality you measure yourself.

The tier question that A​libaba has not resolved

Anyone choosing Qwen-Audio-3.0-TTS today is choosing within a family that moved underneath it. The Qwen-Audio-3.1 line was announced on September 23, 2026, and Artificial Analysis lists a Qwen-Audio-3.1-TTS-Plus at 1,300 Elo on Provider Voice and 1,186 on Controlled — above the 3.0 Plus on both boards, dated September 24, at the same $19.31 per million characters.

But A​libaba's public rate card lists only a Flash SKU for 3.1 T​TS, so whether the 3.1 Plus tier is callable today is a question for the vendor's own console rather than something to infer from a leaderboard. Either way the practical reading is the same: the 3.0 Plus tier is not obsolete, but it is no longer the newest thing A​libaba has scored, and a buyer should confirm which SKU they are actually contracting before comparing its price to anything.

For Ming-3.0-tts-tiny there is no tier question at all — there is one checkpoint, no roadmap with dates, and a vLLM-Omni support note that appeared in the documentation on the same day the code did.

What the repositories do not establish

Four gaps are worth naming before anyone plans around this model.

• No independent score — no arena row, no third-party word error rate, no measured throughput. The row that looks like Ming-3.0-tts-tiny's belongs to the 0.5B.

• No technical report — the roadmap lists it unchecked, and the model card was edited on October 16 to remove a reference to one.

• No hosted endpoint — there is no first-party API and no price, so the only path to production is your own infrastructure.

• No fine-tuning recipe — the repository documents inference tasks, not training. LoRA or full fine-tuning on this checkpoint is undocumented, which is a different situation from a model that ships training scripts.

At the time of the capture the Hugging Face repository showed 20 downloads and 3 likes, and the GitHub repository a handful of stars and no releases. That is not evidence of anything except that the release was quiet, which is the one claim this piece is confident about.

What this comparison is worth in practice

If you need a voice in production this month, in more than a couple of languages, with someone else carrying the uptime, Qwen-Audio-3.0-TTS-Plus is the defensible choice: it is scored, priced, documented and supported, and the 1,266 Elo is measured rather than asserted.

If you need to own the model — clone voices you cannot send to a third party, generate at a volume where per-character billing dominates, ship under a licence that permits redistribution — Ming-3.0-tts-tiny is the only one of the two that can be yours. Accept in exchange that you will be the first party to benchmark it, that Korean is a known weakness its own vendor publishes, and that the pipeline is a week old.

OrcaRouter does not route either model, and nothing here should be read as claiming otherwise; speech is reached through each vendor's own channel, and the Ming-3.0-tts-tiny checkpoint through your own hardware. What we do route is the text half of the same pipeline — more than 200 models behind one key, at provider list price with no markup added, so a price change from a text provider is your price the same day. For a voice product the script, the SSML-like direction and the agent that decides what to say are all text, and that is the part where a second vendor contract buys you nothing.

The number that would settle this does not exist yet. It is not an Elo score; it is one afternoon on an H800 running the tiny checkpoint against the same reference clips Ant used, with the results published alongside the harness. Until someone does that, the honest position is the one this piece has taken: two credible artefacts, one scored and one not, and a comparison that has to be made on ownership, price shape and language risk rather than on quality.