Fish Audio Drama 3 vs Ming 3.0 TTS: The 91-Point Lead That Disappears Under a Cloned Voice

Fish Audio Drama 3 leads Ming 3.0 TTS by 91 Elo points on the leaderboard where each model speaks in its own cast of voices, and by 17 on the leaderboard where both are handed the same eight cloned references — a gap narrow enough that their confidence intervals overlap. Both are open-weight releases from the last six weeks, both are scored by Artificial Analysis, and on the one test that asks a human whether a hard word actually came out of the mouth right, they sit at 93.2%. Three numbers, two of them pointing different ways. Which one you believe decides which model you buy, and the answer depends on whether your product sounds like itself or sounds like somebody else.

Two releases, six weeks apart, built to opposite scales

Drama 3 was announced on 13 October 2026 and is live on Fish Audio's API as drama-3, with weights on Hugging Face under a new community licence. Ming 3.0 TTS arrived more quietly and in pieces: the InclusionAI Hugging Face repositories appeared on 22 September, the 0.5B weights on 24 September, and the inference code only on 19 October — which is why Artificial Analysis could not score it until 20 October. Ant Group has never issued a launch post for it, and it has no hosted endpoint or list price.

The spec sheets are not close, and that is the first thing to understand before the leaderboards mean anything.

• Parameters — 35B mixture of experts with roughly 3B active per token, initialised from Qwen3.5-35B-A3B and routing 8 of 256 experts plus one shared expert, against a 0.5B dense model on a Qwen2.5-0.5B backbone. Ant also ships a 7.9B-total / 1.3B-active MoE called Ming-3.0-tts-tiny, which has no independent score at all.

• Language coverage — 83 languages claimed for Drama 3, against seven for the 0.5B Ming checkpoint (Chinese, English, German, Spanish, French, Korean, Japanese) plus Sichuanese and Cantonese when you supply a dialect reference clip.

• Output — 44.1 kHz from both, generated through very different machinery: Drama 3 keeps its predecessor's Descript Audio Codec tokenizer at ten codebooks and about 21.5 frames per second, while Ming decodes a continuous latent through an AudioVAE.

• Weights and licence — Drama 3 under the Fish Audio Community License v1.0, free below US$500,000 in annual revenue and negotiable above it or if you resell synthesis as a service. Ming 3.0 TTS under MIT, with no revenue gate at all.

• Price — US$50 per million UTF-8 bytes at list, discounted to US$10 until 12 November 2026, against no price because there is nothing to buy: you download the 2.96 GB checkpoint and run it.

The scale difference cuts both ways. Drama 3's 35B backbone is what lets it hold a direction across a long script and cast several speakers into a single scene, with one voice ID per speaker passed as an array; it is also why nobody is going to put it on a laptop. Ming's 0.5B is a model you can serve on a single modest GPU and fine-tune on a rented one.

Why Artificial Analysis gives two different answers

The board everyone quotes first is the Provider Voice Arena, and there Drama 3 has opened a real lead. As of 29 October it sits sixth of 100 models with an Elo of 1,278, priced by Artificial Analysis at US$15 per million characters under its default settings. Ming 3.0 TTS is twenty-fourth at 1,187, with no price recorded because there is no commercial endpoint to record. A 91-point gap on a board whose confidence intervals run around ±13 to ±18 is not noise; Drama 3's native voices are preferred, and clearly. One detail from the same table explains part of why: Artificial Analysis scored Ming against three arena voices and Drama 3 against eight.

The second board asks a different question, and the lead evaporates. The Controlled Voice Arena hands every system the same eight cloned voices — four US, four UK — so the comparison stops being about whose voice actors are better and starts being about how well each model wears somebody else's voice. There Drama 3 scores 1,085, twelfth of 49, with an interval of 1,069 to 1,101. Ming scores 1,068, seventeenth, with an interval of 1,053 to 1,083. The intervals share the range 1,069 to 1,083. That is a tie, and saying otherwise would be reading precision into a survey that does not have it.

Note also that Artificial Analysis still carries a separate Drama 3 Preview row on that board at 1,063, ranked twentieth. The production checkpoint and the preview that shipped in September are scored as two different systems, so a comparison that cites "Drama 3" without saying which row it means is not a comparison. Our numbers here are the production row.

The practical reading: if your product ships its own voice cast — a narrator, a support agent, an in-game character — Drama 3's advantage is genuine and large. If your product clones a customer's or a creator's voice and speaks in that, the two models are effectively level, and the decision falls to price, licence and language coverage instead.

The 93.2% pronunciation tie is four different results

Both models are quoted at 93.2% on Artificial Analysis' English pronunciation robustness test — the result the board published on 20 October, which put them jointly first of 31 systems and ahead of Eleven v4 at 91.7%. That headline has been repeated everywhere, and it hides the only part that matters operationally: the two models fail on opposite kinds of text.

• Exact sequences — Ming preserves them at 90.6%, first of 31; Drama 3's equivalent figure is 86.3%. Order numbers, URLs, file paths and serial codes survive Ming better.

• Shorthand expansion — Ming 96.5%, also first of 31, against 93.6% for Drama 3. Dates, units and abbreviations said aloud rather than spelled out.

• Standalone terms — Drama 3 at 95.9% against Ming at 88.5%, where Ming ranks twenty-first of 31. Brand names, place names and technical vocabulary are Drama 3's strongest category.

• Contextually appropriate terms — Drama 3 at 95.3%, Ming at 91.0% and thirteenth of 31.

Ming's weakness on standalone names tracks its voice inventory: the 0.5B ships more than a hundred built-in Chinese voices and five English ones, so its English proper-noun handling has less to work with. Drama 3's weakness is the mirror image — a model trained to deliver a performance will sometimes deliver a code number as a performance.

Neither 93.2% is vendor-reported. Both come from the same independent harness, which is why they can be compared at all — and it is worth saying that this is rare in this category, where most published word-error tables are run by the labs on their own prompts.

What a million characters actually costs

Drama 3 bills by input size: US$50 per million UTF-8 bytes, cut to US$10 until 12 November 2026 at 09:00 Pacific. That five-fold jump on 13 November is the single largest number in this comparison, and it changes the ranking of the two models on cost per finished hour depending on when you read it.

There is also a billing trap worth knowing before you wire anything up. Drama 3 keeps the same POST /v1/tts endpoint as S2.1 Pro, and if the model header is missing or misspelled the request silently falls back to s2.1-pro and bills you for it. You get plausible audio and no error, so a typo in an environment variable can run for weeks before anyone notices the voice is wrong.

Ming has no invoice at all, which is not the same as being free. You pay in GPU hours and in engineering time: the 0.5B checkpoint is 2.96 GB, needs the audio VAE, a CAM++ speaker encoder and a tokenizer wired together, and has no vendor support commitment behind it. Ant has published benchmark tables and a Gradio demo, not a service-level agreement.

This is where a router is genuinely useful and where it is genuinely not. OrcaRouter passes provider list price through at 0% markup, so a vendor promotion like Fish Audio's lands the same day on anything we route — Drama 3 is not one of them, and neither is Ming. The speech endpoints we do carry today are Google's Gemini TTS line and OpenAI's TTS-1-HD, all reachable through one API behind a single key alongside the text models. If you are comparing Drama 3 and Ming, no routing layer is going to host either for you; what a single key buys is the rest of the pipeline around them.

The licence decides more than the leaderboard does

MIT versus the Fish Audio Community License v1.0 is the sharpest fork in this pair, and it is not a philosophical one. Ming's MIT terms let any company of any size embed the model in a paid product, resell it, fine-tune it and ship the derivative, with no revenue threshold and no conversation with Ant.

Drama 3's licence is free for research and non-commercial work, free for commercial use while you and your affiliates stay under US$500,000 in annual revenue, and requires a negotiated licence above that — or at any size if you want to offer Drama 3 as a text-to-speech, cloning or inference service. The model card's own FAQ is clearer than most: finished audio for clients is fine, a competing synthesis API is not.

One detail to check rather than assume: the Hugging Face metadata for fishaudio/drama-3 still reads fish-audio-research-license while the card text describes the Community License. If you are the person signing off on the deployment, read the licence file in the repo, not the tag.

For a studio making audio drama, the $500,000 gate is a non-issue and the service clause is the one to read. For a company building a voice feature into a product above that revenue line, Drama 3 is a procurement conversation and Ming is a pip install.

Which one to pick

Support or transactional voice, English or Chinese. Ming. It wins the two categories that matter most when a bot reads back an order number or a booking reference, it is MIT, and it runs on hardware you already have. The cloned-voice tie means you are not giving up naturalness to get there.

Narration, audio drama, brand-heavy scripts, many languages. Drama 3. The provider-voice lead is real, proper nouns are its strongest category, and 83 languages versus seven is not a gap you close with prompt engineering. Budget for the November price change and read the service clause.

A pipeline that already does multiple takes. Drama 3, but only if you actually use the revision loop. Fish Audio trained it with a human-preference reward model and pitches multi-turn revision as the way to get the performance you wanted; Ming is a single-shot model with no revision mechanism to weigh against it, so the take structure is a reason to choose Drama 3 rather than a reason to compare them.

Any deployment where VRAM is the constraint. Ming, without argument. A 35B mixture of experts that benchmarks on H200s is not a model you run because it is convenient.

What to watch

Three dates and one gap. On 13 November the Drama 3 promo price expires and US$10 becomes US$50, a five-fold move that will reorder every cost comparison in this category overnight. Before that, watch whether Artificial Analysis merges its Drama 3 and Drama 3 Preview rows, because a board that lists one model twice invites exactly the kind of sloppy citation this article is trying not to make. And watch for a score on Ming-3.0-tts-tiny: the 7.9B MoE variant is the one Ant actually pitches for podcast generation, and right now it has only vendor numbers — 0.820% word error on Chinese, 6.000% on Korean — with nothing independent behind them.

The honest summary is that these are not two versions of the same product. One is a small, permissively licensed workhorse that happens to be excellent at exactly the text that breaks voice systems; the other is a large, licence-gated performance engine that happens to be the best-sounding open model on the board when it uses its own cast. The 91-point lead is real. It just belongs to a question most products never ask.