GPT-Live-1 vs Ming-3.0-TTS: An MIT Licence Meets a Metered Minute

Ming-3.0-TTS arrived on 20 October 2026 with the licence most voice teams dream about: MIT, weights on Hugging Face and ModelScope, inference code on GitHub, zero cost per character, and Sichuanese and Cantonese in the language list. GPT-Live-1 costs $0.05 every minute, ships no weights, and speaks the languages OpenAI tuned it for. Comparing them is not really a quality question. It is a question about which language your customers actually speak and whether you are willing to run a GPU to serve them.

Both are real, both are current, and both are good at things the other cannot do. Here is the honest split.

The Ming family is two models, not one

Ant Group's inclusionAI team open-sourced Ming-3.0-TTS as a pair sharing one generation framework — a language model feeding a diffusion transformer patch head and an audio tokenizer.

• Ming-3.0-TTS-0.5B — dense, roughly 500 million parameters on a Qwen2.5-0.5B backbone, a 2.96GB download, runs on an H800-80GB or an H20-96G.

• Ming-3.0-TTS-Tiny — a mixture-of-experts build on Ant's Ling V3 architecture, 7.9 billion parameters total with 1.3 billion active, and the one that adds a multi-speaker podcast mode.

The 0.5B is the interesting one for anyone who has priced a voice product. A half-billion-parameter TTS model that fits in three gigabytes is not a data-centre commitment; it is a single GPU you might already have. The Tiny is the quality play, and Ant's own notes say its podcast output is below the earlier Ming-Omni-TTS — a candid limitation, and one worth taking at face value.

Feature set: zero-shot voice cloning, voice design from a text description, more than 100 built-in Chinese voices and 5 English ones, 14 inline paralinguistic tags for laughs and sighs and breath, natural-language style control, and 44.1kHz output. Languages are Chinese, English, German, Spanish, French, Korean and Japanese, plus Sichuanese and Cantonese dialects.

What GPT-Live-1 offers instead

OpenAI's model is a conversation engine, not a synthesiser with a script interface. On the API since 10 September 2026, it takes audio and text in, returns audio and text out, listens while it speaks, and pushes reasoning to a text model when a question needs one. One endpoint: v1/live/sessions. $0.05 per minute, billed by the second. Knowledge cutoff July 2025, no documented context window, and no Free-tier access.

There is no documented dialect support, no voice cloning, no per-character billing, and no way to hand it a 4,000-character script and get a file back. What it does have is turn-taking that works — the thing every cascaded voice agent spends six months tuning and still gets wrong on a bad phone line.

The language question comes first

If your users speak Mandarin, Cantonese or Sichuanese, this comparison ends early. Ming-3.0-TTS was built for those languages, publishes per-language word-error rates, and ships voices that sound native rather than adapted. GPT-Live-1's documentation does not describe dialect handling at all, and the deployments we can observe — ChatGPT Voice, and now Perplexity's voice mode — are English-first products.

If your users speak English only, Ming's advantage shrinks to its licence and its price, and GPT-Live-1's advantage becomes the entire conversational layer. That is the real fork, and it has nothing to do with Elo.

The one place Ming's language breadth is genuinely strategic is a product that must serve several markets at once — German and Spanish and Korean and Japanese in one release. Ant reports its best German and Spanish word-error rates in the Tiny variant, while noting speaker similarity is weaker in newer languages and mostly sits outside the top three. Translation: broad coverage, uneven fidelity, and you should audition per market rather than trust the list.

The $0 claim, audited

Ming is free in the sense that the licence costs nothing. It is not free in the sense that serving it costs nothing, and every "zero-cost TTS" write-up quietly skips that step.

Artificial Analysis records no price for Ming-3.0-TTS at all — the field reads N/A, because there is no hosted commercial API from inclusionAI to price. Your cost is GPU time, and the honest formula is: (GPU hourly rate ÷ characters synthesised per hour) × 1,000,000. Fill in the second term yourself. Nobody publishes a throughput number for the 0.5B, so any per-million figure you read online — including the $6.58 break-even floating around in comparisons — rests on an assumption about speed that the model's authors have not published.

What can be said without inventing numbers: at a $2 per GPU-hour rental, you need to synthesise roughly half a million characters per hour to land under $4 per million — a rate that requires the model to run several times faster than real time, which a 0.5B model plausibly can and a 7.9B MoE plausibly cannot. Measure it before you build a business case on it.

Against GPT-Live-1 the arithmetic is cleaner precisely because the meter is linear: one hour of conversation is $3.00, with backend reasoning billed separately. No utilisation risk, no GPU to babysit, and no assumption you have to defend in a budget meeting.

Why both report 93.2%

A curiosity that will trip up anyone comparing datasheets: Ming-3.0-TTS reports 93.2% pronunciation robustness on Artificial Analysis' measurement, claiming first overall. Fish Audio's Drama 3 reports the same 93.2% and also claims first. Two labs, two architectures, one identical figure to a decimal place.

Treat that as evidence of a ceiling rather than a tiebreak. The remaining fraction is concentrated in the cases that break every TTS model: alphanumeric strings, unusual proper nouns, numbers read in context, and homographs. Which model wins for you depends on whether your content contains part numbers or poetry, and only your own eval set can answer that.

On the boards, with the caveats attached

On the Provider Voice Arena captured 29 October 2026, Ming-3.0-TTS sits at 1187 Elo, third among open-weights entries behind Drama 3 at 1240 and Breeze TTS 2 at 1228, and the ranking Ant's own repository reports as third among open-weights on this board. On the Controlled Voice board, where every model clones the same eight voices, Ant's own repository claims first among open-weights at 1069 Elo and thirteenth overall.

GPT-Live-1 does not appear on those boards, because it is not a text-to-speech model in their sense. Its scores live elsewhere: 83.2 on the Speech-to-Speech Index, second of 21, and 1037–1041 on the Speech Agent Arena with task success between 87.4% and 90.9%. Different instruments, different scales, no subtraction permitted.

Where each one loses

Ming loses on conversation, full stop. It cannot hear, cannot interrupt, cannot decide to stop talking. Any deployment that needs a human on the other end of the line is out of scope, however good the voice is.

Ming also loses on operational simplicity for small teams. A GPU is a thing that goes down at 3am, needs drivers that match, and quietly degrades when another tenant shares the card. The MIT licence does not include an on-call engineer.

GPT-Live-1 loses on cost structure for text-heavy work and on control for brand voice. Fixed model voices mean you cannot ship a founder's voice or a licensed character. And the per-minute meter punishes exactly the workload TTS is best at: long, scripted, repetitive output where nothing needs to listen.

The routing angle that actually applies

Neither of these is on OrcaRouter, and it is worth being blunt about it: Ming-3.0-TTS is self-hosted open weights with no hosted endpoint to route, and GPT-Live-1 exposes a single session endpoint we do not serve. What we do route is the reasoning model behind a Chinese-language voice agent — Qwen3.8 Max, GLM-5.3, Kimi K3 and Gemini models sit behind one key with automatic failover, at provider list prices passed through with 0% markup.

For a bilingual deployment that matters more than it sounds. The TTS choice is usually made once, on language coverage; the reasoning model behind it gets swapped as prices move and new Chinese-language checkpoints land. Keeping that leg routable is the cheapest form of future-proofing in the stack.

Which one to build on

Build on Ming-3.0-TTS if the product speaks Chinese or serves multiple Asian and European markets, if you control a GPU, and if the output is scripted rather than conversational. The MIT licence is a genuine strategic asset: it cannot be revoked, repriced, or rate-limited, and the 0.5B variant makes the barrier to trying it roughly one afternoon.

Build on GPT-Live-1 if the product is a conversation in English and the quality bar is "does it feel like a person on the phone". You will pay $0.05 a minute forever and own nothing, and you will get the one capability the open-weights field has not shipped at this quality: a model that listens while it speaks.

Do not build on either because a leaderboard said so. Ming's 1187 and Live's 83.2 are two different instruments pointed at two different jobs, and the only comparison that survives contact with a real deployment is the one you run on your own audio.