Ming-3.0-tts vs ElevenLabs: Eleven v4 Wins the Arena by 140 Points and Loses the One Test That Decides Whether Your Text Gets Read Correctly

Ask a room of voice-AI engineers to pick between Ming-3.0-tts and Eleven v4 and you will get ElevenLabs almost every time, and the arena agrees with them: 1,328 Elo for Eleven v4 on Provider Voice against 1,188 for Ant Group's model, a 140-point gap that is not a rounding error, a tie, or a temporary artifact of vote counts. Then open the pronunciation board, where the same lab asks human reviewers whether each highlighted span in an English sentence was said correctly, and Ming-3.0-tts sits at 93.2% against Eleven v4's 91.7% — first place against third. A model that sounds noticeably worse to listeners is also, by an independent measure, more reliable at saying the words that must be said exactly. 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.

That is not a contradiction, and understanding why it is not is the entire value of this comparison. ElevenLabs sells a voice. Ming-3.0-tts is a checkpoint you run. The overlap where they genuinely compete is narrower than either vendor's marketing implies, and inside that overlap the answer changes with the kind of text you feed them.

What ElevenLabs actually ships

ElevenLabs now runs two current model lines. Eleven v4 is the emotive, high-fidelity one: 90-plus languages, a 10,000-character limit per request, multi-speaker dialogue, and voice cloning that starts at the $6 Starter tier for instant clones and the $22 Creator tier for professional ones. Eleven v4 Turbo is the real-time sibling, same 90-plus languages, with a published median inference latency of about 100 milliseconds excluding application and network overhead, and API pricing at roughly half of v4's. Pricing runs on credits at one credit per character, with plans from free at 10,000 credits to $990 per month for six million; the arena normalises v4 to $80 per million characters and Turbo to $40.

Ming-3.0-tts is a different object entirely: a 2.96 GB MIT-licensed checkpoint from Ant Group's inclusionAI team, built on a Qwen2.5-0.5B backbone with a diffusion-transformer patch head, released as weights on 24 September 2026 and made runnable when the inference code landed on GitHub on 19 October. It covers seven languages plus Sichuanese and Cantonese, clones a voice from reference audio plus its transcript, designs voices from text descriptions, and takes 14 inline paralinguistic tags. There is no hosted endpoint, no subscription, and no per-character meter.

ElevenLabs' second model line matters here and rarely gets mentioned in comparisons: Eleven v3, at 1,072 Elo, still in production, and v3 Conversational at 1,062 for real-time use. A team that picks "ElevenLabs" is often picking v3 because it is cheaper or because their integration predates v4. Both score below Ming-3.0-tts on the preference board. The comparison is with v4, and v4 is a genuinely different tier from the v3 line most existing deployments run.

The preference gap is real, and it is not where you think

• Provider Voice Elo — Ming-3.0-tts 1,188 (1,174–1,202) vs Eleven v4 1,328 (1,314–1,342). Eleven v4 is 2nd of 97; Ming is 18th. Eleven v4 Turbo leads the board at 1,330.

• Controlled Voice Elo, same eight cloned voices for every system — Ming-3.0-tts 1,068 (1,052–1,084) vs Eleven v4 1,156 (1,143–1,169) and Turbo 1,158. The gap narrows from 140 to 88 points.

Read the second line again. When both systems clone the same voice, over half the ElevenLabs advantage evaporates. Provider Voice measures the whole product — the native voice catalogue, the default delivery, the prosody choices a vendor made for you — and ElevenLabs has spent years on that catalogue while Ming ships five English voices against a hundred-plus in Chinese. Controlled Voice strips that away and asks only how well each system imitates a voice it has been handed. ElevenLabs still wins, comfortably, but the margin is a different size and it is measuring a different competence.

For a team cloning its own brand voice, the 88-point number is the relevant one. For a team that wants to pick a voice from a catalogue and ship, the 140-point number is, and it reflects an advantage Ming cannot close without building a voice library it has shown no sign of building.

Where the ranking inverts

The pronunciation board weights two categories heavily, and Ming-3.0-tts wins both outright.

• Preserving Exact Sequences — codes, paths, emails, identifiers. Ming-3.0-tts 90.6%, first of 31, against Eleven v4 at 78.8% and Eleven v4 Turbo at 73.9%. That is a 12-point gap on the category that decides whether an order number, a repository path or a customer email address comes out intact.

• Expanding Shorthand — numbers, dates, units, notation. Ming 96.5%, first of 31, against Eleven v4 at 94.1% and Turbo at 91.9%.

And where ElevenLabs wins, it wins cleanly:

• Standalone Terms — brand, place and technical names. Eleven v4 93.2%, Ming 88.5%. Ming ranks 21st of 31 here, which is the single most important line in this article for anyone producing English marketing or support audio.

• Contextually Appropriate — heteronyms such as "wound." Eleven v4 94.1%, Ming 91.0%.

So the split inside one test is: Ming reads your data correctly and mispronounces your brand; ElevenLabs reads your brand correctly and mangles your identifiers. Both failures are audible and neither shows up in a preference vote, because a listener who does not know the correct pronunciation of a term has no way to mark it wrong.

The aggregate — 93.2 against 91.7 — is a tie in practice. The confidence intervals overlap (91.5–94.9 for Ming, 90.0–93.4 for v4), the sample sizes differ (1,800 spans against 2,051), and a 1.5-point lead on a 31-model board is not a claim worth building a decision on. What is worth building on is the category split, because there the intervals separate.

Cost, converted honestly

ElevenLabs bills one credit per character, with v4 at the standard rate and Turbo and Flash variants at half to one credit depending on plan. On the published plan table, the effective per-minute cost of T​TS lands around $0.17–$0.36 depending on tier, and the arena's normalised figure is $80 per million characters for v4 and $40 for Turbo.

Worked against real volume: a million characters is roughly 20 hours of finished speech at a normal narration pace. At $80 that is $4 per finished hour of audio, which is cheap until it is not — a catalogue of 5,000 product descriptions at 600 characters each is three million characters, or $240 on v4 and $120 on Turbo. That is a monthly bill for a content operation that regenerates its copy.

Ming-3.0-tts has no per-character price because there is no hosted service. The cost is GPU time at the measured 44.3 characters per second, which puts a million characters at about 6.3 hours of generation on an H800-class card, plus loading. Whether that beats $80 depends entirely on what you already pay for compute and how much engineering time the self-hosted path consumes. For a single product description a month, ElevenLabs is cheaper by a factor that makes the comparison silly. For a nightly regeneration of a large corpus, the arithmetic changes.

There is a third option that neither vendor's pricing page will suggest, and it is where a routing layer earns its keep. OrcaRouter does not host Ming-3.0-tts and does not carry ElevenLabs, so neither side of this matchup is reachable through us; Ming comes from Ant's repository and ElevenLabs from its own API. What we do carry is G​oogle's G​emini T​TS previews and O​penAI's tts-1 and gpt-4o-mini-tts alongside more than 200 text models, all at provider list price with no markup added, with automatic failover across providers behind one key. If the voice is one stage in a pipeline whose earlier stages are text — summarisation, translation, script generation — that is the part that consolidates today, and a provider price change lands the same day rather than at the next contract renewal.

The engineering difference that decides more than the Elo

ElevenLabs gives you a request, a response, a concurrency limit that scales from 2 on the free tier to 15 on Scale, and a vendor who answers when it breaks. Ming-3.0-tts gives you a 2.96 GB checkpoint, a Docker image using PyTorch 2.6 and CUDA 12.4, a Python class, and Ant's evaluation tables. The repository's roadmap lists vLLM-Omni accelerated inference as complete, but the technical report as still unstarted.

Two operational facts follow, and they are not small.

First, Ming-3.0-tts requires a reference waveform and its transcript for cloning — the --prompt_wav_path and --prompt_text arguments in the repository's own example. ElevenLabs clones from ten seconds of audio on Starter and above, with no transcript and no local pipeline. If your cloning workflow is "a customer uploads a voice memo," that difference is a product decision rather than a technical detail.

Second, capabilities in Ming are task modes, not composable switches. Ant's own documentation says the abilities are exposed as separate tasks and cannot be freely combined, which means a request that wants cloning plus emotion plus a dialect is not necessarily a request the model accepts as one call. ElevenLabs exposes emotion and audio tags in the same request as the cloned voice. For an agent that needs all three at once, this is the constraint that matters most and it does not appear on any leaderboard.

A note on which ElevenLabs number to trust

ElevenLabs' pricing page promotes v4 as its newest and most emotive model while its comparison table still lists v3, and the arena separately scores v3, v3 Conversational, v4 and v4 Turbo as four distinct systems with a 260-point spread between the top and bottom. "ElevenLabs" is not a single competitor in this comparison; it is a product line, and the gap between its best and worst current model is larger than the gap between Ming-3.0-tts and either of them. Any comparison that uses one number for ElevenLabs is choosing a number that flatters or penalises the vendor depending on which one it picked.

What to do, concretely

If the audio is customer-facing conversation — a phone agent, an interactive assistant, anything where the first syllable arrives while someone is waiting — use Eleven v4 Turbo. The latency claim is published, the concurrency tiers are published, and Ming-3.0-tts has no streaming path at all.

If the audio is English marketing, brand-led narration, or anything where a proper noun is the point of the sentence, use Eleven v4 and budget for the pronunciation dictionary. Ming's 88.5% on standalone terms is 21st of 31 and it will be audible on the first brand name.

If the audio is operational — order confirmations, technical read-aloud, internal training, anything dense with identifiers, units and notation — Ming-3.0-tts is the better instrument and it is not especially close on the two categories that cover those spans. It is also the only one of the two you can keep if a vendor changes its terms, because MIT does not expire.

What would change the answer is the same thing it has been for a month: Ant still has not published a technical report, and Ming still has no hosted endpoint. Once either lands, the 88-point controlled-voice gap is the number to re-measure, because that is the one closest to what a team cloning its own voice will actually experience.