Ming-3.0-tts-0.5B vs Qwen-Audio-3.0-TTS: Free Weights Against 78 Elo Points

Put Ming-3.0-tts-0.5B next to Qwen-Audio-3.0-TTS and the first thing you notice is that they are not the same kind of object. Ming is a 2.96 GB checkpoint you download, licence under MIT and run on your own GPU. Qwen-Audio-3.0-TTS, in the quality tier A​libaba lists as Qwen-Audio-3.0-TTS-Plus, is a hosted endpoint on A​libaba Cloud Model Studio with no weights anywhere. On the independent Artificial Analysis speech arena the hosted model scores 1,266 Elo and the downloadable one scores 1,188 — a 78-point gap with no overlap in the two confidence intervals. That gap is real. What it is worth depends on whether you are buying a voice or buying an asset. 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 short answer, before the evidence

If the voice ships in a product this quarter and someone else should be on call when it breaks, Qwen-Audio-3.0-TTS-Plus is the better pick and it is not especially close. If you are building Chinese-first content, need voice design or dialect control, or cannot accept a per-character meter, Ming-3.0-tts-0.5B is the only one of the two that can be yours. Neither answer survives contact with a budget spreadsheet unchanged, which is what the rest of this piece is about.

What each model actually is

Ming-3.0-tts-0.5B is Ant Group's open-weights release from inclusionAI, built on a Qwen2.5-0.5B backbone with a diffusion-transformer head and a separate audio tokenizer. It shipped quietly in two stages — weights on Hugging Face in September 2026, inference code on GitHub on October 19 — and Ant has not announced it. The bundle includes a speaker encoder and a text tokenizer, outputs 44.1 kHz audio, and supports cloning, voice design from a text description, 14 inline paralinguistic tags, emotion and attribute control, seven languages and two Chinese dialects.

Qwen-Audio-3.0-TTS is A​libaba's hosted synthesis line, launched July 20, 2026, in two tiers: Flash for real-time work and Plus for maximum quality. The Plus tier is the one with a measured position on the arena, and it is the tier every number below refers to. It covers 16 languages plus 20 Chinese dialect regions, supports voice cloning and roughly 86 inline delivery tags, and is reachable only through A​libaba's API. There are no weights, no local deployment and no licence to audit.

One dating oddity worth flagging because you will see both dates: A​libaba launched the line in July, but the arena row for the Plus model carries a release date of September 15, 2026. We read that as the date the scored revision or listing went up, not a second model. If you are citing a release date in a document, July 20 is the launch and September 15 is the arena's bookkeeping.

The numbers that separate them

• Arena Elo, Provider Voice — Ming-3.0-tts-0.5B 1,188 (interval 1,173–1,203) vs Qwen-Audio-3.0-TTS-Plus 1,266 (1,252–1,280). The intervals do not touch.

• Arena Elo, Controlled Voice — Ming 1,069 (1,053–1,085) vs Qwen-Plus 1,126 (1,113–1,139). Still a clear gap, 57 points, on cloned voices held constant across all models.

• Price — Ming has none; you pay for compute. Qwen-Plus is listed at $19.31 per million characters on the arena's normalised price column.

• Throughput — Ming 42.1 characters per second of generation time vs Qwen-Plus 25.6, measured on the same independent harness.

• Weights — MIT-licensed checkpoint vs none.

• Languages — seven plus Sichuanese and Cantonese vs 16 plus 20 Chinese dialect regions.

• Control surface — free-form instruction, voice design, 14 paralinguistic tags, explicit emotion/speed/pitch/volume vs roughly 86 inline tags and instruction-style delivery control.

The throughput result is the one that surprises people. A half-billion-parameter dense model with a 2.96 GB bundle outruns A​libaba's hosted Plus tier by about 65% on characters per second in the independent measurement. That is not a knock on Q​wen — hosted endpoints are provisioned for concurrency, and the arena's throughput number is one request at a time — but it does mean the small model is not the slow one in this comparison.

Where the 78 points come from

Preference Elo does not tell you what is wrong. Listening to the two systems on the same text does, and the pattern is consistent with what Ant's own model card admits.

Ant publishes its weaknesses in unusually plain language: speaker similarity trails content accuracy, French similarity is the clearest multilingual gap, multi-event paralinguistic sentences are still unstable, and open-ended voice design does not yet match the strongest system evaluated. On the arena, that shows up as a cloned voice that is competitive — first among open-weights models on the controlled board — and a native voice cast that is not. Ming's provider-voice rank, 19th of 98, is dragged down by the 100-plus Chinese and five English built-in voices rather than by the cloning path.

Q​wen's advantage is the opposite shape. Its native voices are the product; the Plus tier was built to sound finished out of the box, and it measures that way in both the provider and controlled settings. It also has the language breadth that comes from a company running a cloud: 16 languages and 20 Chinese dialect regions against Ming's seven and two.

On accuracy rather than preference, the comparison is thinner than the Elo gap suggests. Qwen-Audio-3.0-TTS has no published word-error table from A​libaba that we could verify against a common test set, and Ming's Seed-TTS numbers — 0.800% Chinese, 1.584% English — are Ant's own, measured on Ant's harness with Ant's recommended settings. Neither side gives you an apples-to-apples error rate. What you have is a preference score for both and a vendor table for one.

Two billing models, and the meter nobody prints

The price comparison is not $19.31 versus zero. It is a per-character meter against a GPU-hour meter, and the second one is invisible until your finance team asks for it.

Q​wen bills per character of input text. At $19.31 per million characters, a million characters of finished narration costs $19.31, and the number scales linearly and predictably. You can put it in a spreadsheet before you write a line of code.

Ming bills in GPU time. The release has been tested on an H800-80GB and an H20-96G, needs PyTorch 2.6 with CUDA 12.4, and ships a Dockerfile but no hosted endpoint. At the arena's measured 42.1 characters per second, a million characters takes about 6.6 hours of generation on that class of hardware, before you count model loading, queueing or failed requests. Whether that is cheaper than $19.31 depends entirely on what your GPU-hour costs and how full you keep it — and it is the calculation that decides this matchup for most teams.

This is also where a routing layer earns its keep without pretending to host either model. OrcaRouter does not carry Ming-3.0-tts-0.5B or Qwen-Audio-3.0-TTS, and nothing here should suggest otherwise; both are reached through their own channels, A​libaba's Model Studio for Q​wen and your own infrastructure for Ming. What we do carry is the text side of the same pipeline at provider list price with 0% markup added — so a price change from a text provider we route is your price the same day — with automatic failover across providers for the model that writes the script, tags the emotion or drives the agent. Speech is the expensive, slow half of a voice product; the half in front of it should not also be the fragile one.

The successor problem

There is a third model that belongs in this decision and is not in the title: Qwen-Audio-3.1-TTS-Plus, released September 24, 2026. It scores 1,300 Elo on the same arena — 34 points above the 3.0 Plus tier — at the same $19.31 per million characters.

Read that again as a buyer. A​libaba shipped a better model five weeks after the one in this comparison, at an unchanged price, on the same endpoint family. Anyone choosing Qwen-Audio-3.0-TTS today is choosing a model its own vendor has already superseded at no discount. The 3.0 tier remains listed and remains good — 1,266 Elo is sixth on a board of 98 — but the correct hosted choice in this family is now 3.1, not 3.0.

Ming has no such problem because it has no successor and no cadence. That is a genuine advantage for a team that wants a stable artefact and a genuine risk for one that wants the best voice available in six months. An unannounced release with no roadmap commitment can go quiet in either direction.

Who should pick which

Pick Qwen-Audio-3.1-TTS-Plus — not 3.0 — if you want the highest measured quality in this family at a known per-character price, need broad language coverage, or want someone else to own the uptime. Accept the meter and accept that the next tier will arrive without warning.

Pick Ming-3.0-tts-0.5B if the voice has to run inside your own perimeter, if your content is Chinese-first and needs Sichuanese or Cantonese, if you want to design a voice from a description rather than pick from a cast, or if your volume is high enough that a GPU you already own beats a per-character bill. Budget for the engineering: the environment is pinned, text normalization needs an FST library that has no universal wheel, and the only published throughput figures come from one independent harness.

Pick neither, for now, if you need a support contract. Ming has no announcement, no hosted option and no stated support commitment; Q​wen has all three, but the 3.0 tier is no longer the one to buy. The honest summary is that this is not a fair fight and was never meant to be — one is a research release with a licence and the other is a cloud product with a rate card — and the useful question is which of those two things your project actually needs.