Ming-3.0-tts vs Sesame Preview: the Voice You Can Own Against the Voice You Can Only Hear

You can hear Sesame Preview in about thirty seconds and you cannot download it at any price. You can download Ming-3.0-tts for free under the MIT license and there is no polished demo to try first. That asymmetry is the entire comparison, and most head-to-heads get it wrong by treating the two as competing text-to-speech APIs. They are not. Sesame Preview is a finished consumer product built around a voice Sesame has never released; Ming-3.0-tts is an open-weight speech model from Ant Group's inclusionAI team, published on 20 October 2026, with no hosted consumer front end at all. One is a restaurant. The other is a kitchen.

That does not make the comparison useless. It makes it a question about what you are actually trying to buy: a conversational agent that works today, or a voice you can put inside your own product and keep. Once you separate those, the decision is clearer than the spec sheets suggest — and it turns on facts that neither model's marketing page states plainly.

Three things called "Sesame," only one of which you can use

The naming is the first trap. "Sesame Preview" refers to the free app with four conversational personas — Maya, Miles, Simone and Charlie. It launched as a web preview in February 2025, reached iOS in late May 2026 and Android on 6 August 2026. It performs live web searches mid-conversation and remembers context across a session.

What powers that app is not the model you can download. Sesame's production voice runs on a larger, closed model that has never been released. The downloadable artifact is CSM-1B, the Conversational Speech Model: about 1B parameters, Apache-2.0, gated behind a terms acceptance, built on a Llama backbone with Kyutai's Mimi audio codec at 24 kHz. CSM-1B is the ancestor of the demo voice, not the demo voice. It is a base model that cannot generate text on its own and needs a language model paired with it.

So the honest framing of "Ming-3.0-tts vs Sesame Preview" is really two comparisons stacked. Against the Preview app, Ming competes on capability and loses on polish and latency. Against CSM-1B, Ming competes directly on weights, license and control, and wins on most of them.

What Ming-3.0-tts brings to the table

Ming-3.0-tts is the speech-synthesis line of Ant's five-model Ming-3.0 family, MIT-licensed, with weights on Hugging Face and ModelScope and inference code in a separate GitHub repository. It ships in two checkpoints that are genuinely different products.

• Ming-3.0-tts-0.5B — dense, on a Qwen2.5-0.5B backbone with a DiT head and an audio tokenizer, about 2.96 GB. This is the checkpoint with independent benchmark runs.

• Ming-3.0-tts-tiny — a 7.9B-total, 1.3B-active mixture of experts on Ant's Ling V3 backbone, roughly 18 GB, adding multi-speaker podcast generation. "Tiny" describes runtime cost, not download size.

Capabilities that matter for this comparison: zero-shot voice cloning from a short reference clip, voice design from a natural-language description, a director mode for free-form delivery control, 14 inline paralinguistic tags such as laughter and sighs, explicit emotion, speed, pitch and volume controls, 100-plus Chinese voices and five English ones, seven languages plus Sichuanese and Cantonese, and 44.1 kHz output.

The voice test neither leaderboard settles

Sesame's reputation rests on how the Preview voices sound. Third-party measurements back that up: CodeSOTA's open-source TTS shortlist placed Sesame first with a 4.7 MOS, and CallSphere measured the production Maya voice at 4.6 MOS. Those are strong numbers, and they are the reason the demo went viral.

They are also measurements of a voice you cannot license. The relevant question for a developer is not "how good is Maya" but "how good is the best voice I can put in my product."

On the independent boards, Ming-3.0-tts is strongest exactly where a cloning workflow lives. Artificial Analysis's Controlled Voice Arena, which uses eight cloned voices across four US and four UK accents, puts Ming at 1,069 Elo — first among open-weight models and around 13th overall. Its Provider Voice score, where a model uses its own stock voices, is 1,188 Elo, third among open-weight models and around 18th overall. The gap between those two results is the model's character in one line: it is much better at reproducing a voice you supply than at being a voice out of the box.

Ming also tied for first overall on Pronunciation Robustness at 93.2%, which is a real result and a narrow one. On standalone terms — brand names, place names, isolated proper nouns — it ranks 21st of 31. Read that as a warning label for anyone whose script contains product names: the aggregate score is not the number your bug reports will be about.

Sesame Preview does not appear on those boards in a comparable form. Its app voices are fixed personas, not clones, and there is no public checkpoint of the production model to benchmark. The 4.7 MOS is the best available evidence and it measures a different thing.

Control: a director's chair versus a personality

Ming treats speech as a controllable signal. You can clone a reference voice, design one from text, tag an utterance with a laugh or a sigh, or instruct the delivery in free-form language. The model card is candid about where this is imperfect: multi-event paralinguistic tags are weaker than single-event ones, emotion control is harder when the text carries little emotional signal, pitch is less accurate than speed or volume, and open-ended voice design still trails the strongest system the team evaluated. None of that is unusual for a first public release. All of it is documented, which is more than most vendors offer.

Sesame's control surface is a conversation. The app manages turn-taking, memory and live search, and the voice is one component of an agent rather than a parameter you tune. Its weakness is the flip side of that integration: function calling has been measured at a 6.7% error rate against 2.1% for a realtime competitor, which is the kind of number that matters if you wanted to build an agent on it rather than talk to one.

If your requirement is "a voice that reads this script the way I specify," Ming is the only one of the two that offers the controls. If your requirement is "a voice agent that handles a conversation," Sesame is the only one of the two that ships one.

Language and audio quality

• Languages — Ming covers Chinese, English, German, Spanish, French, Korean and Japanese, plus Sichuanese and Cantonese. Sesame Preview is English-first; the app is English-only, and CSM-1B's non-English performance is largely untested.

• Sample rate — Ming outputs 44.1 kHz. CSM-1B outputs 24 kHz through the Mimi codec, which is adequate for conversational speech and noticeably narrow for music-adjacent or broadcast work.

• Voice catalogue — Ming ships 100-plus Chinese voices and five English ones. Sesame ships four personas, each with a distinct character, which is a curated product decision rather than a catalogue.

• Emotional range — Ming exposes emotion as a control with documented limits. Sesame's emotional expressiveness is baked into the personas and is widely regarded as its strongest quality; it is also not something you can dial.

Latency is Sesame's unarguable win

This is the one axis where the comparison is not close. Sesame Preview streams with sub-300 ms first-byte latency; the Maya voice has been measured at 80–130 ms time-to-first-byte, and the full agent at 250–450 ms with a 480 ms p95. That is conversational-grade, and it is the result of a product built end-to-end for real-time dialogue.

Ming-3.0-tts is research-grade inference code. There is no documented streaming path, no serving stack, and no latency figure in the model card. The repository points to vLLM-Omni accelerated inference as roadmap. Until that exists, comparing its latency to Sesame's is comparing a benchmark harness to a shipped product, and the shipped product wins by default.

What each one costs

• Sesame Preview — the app is free. The demo voice is not for sale at any price, which is the relevant fact for a developer.

• CSM-1B — Sesame's own API lists it at about $7 per million characters with a 4K context, served by Sesame alone. At roughly 14 characters per second of speech, that is around $0.35 per audio-hour, and it is the cheapest metered speech option in this corner of the market.

• Ming-3.0-tts — free MIT weights and a GPU bill. A rented H800 at around $2 per GPU-hour running a 0.5B model above real time lands well under the metered rate at sustained volume, and above it at low volume once idle capacity is counted honestly.

The cost comparison therefore depends entirely on volume and on whether you already own GPUs. Below a few dozen hours of audio a month, CSM-1B's $0.35 per audio-hour is hard to beat and requires no infrastructure. Above a few hundred hours, self-hosting Ming is cheaper and keeps getting cheaper, because the marginal generation costs nothing. The mistake is choosing on sticker price instead of utilization.

Licensing is where the two actually separate

CSM-1B is Apache-2.0, which is a permissive license, but it is gated behind a terms acceptance and it is not the model in the demo. Ming-3.0-tts is MIT on both checkpoints, ungated, with no acceptable-use click-through and no commercial carve-out. For a legal review that asks whether the model can be fine-tuned on customer voice data and shipped inside a closed product, both answers are yes — but only one of them is the model you have actually heard.

That is the trade in a sentence: Sesame gives you the better-sounding voice and keeps it; Ming gives you a good-sounding voice you can take home.

Where OrcaRouter fits — and where it does not

Neither model is hosted on OrcaRouter. There is no Ming speech endpoint and no Sesame endpoint on our platform, and this page is not going to invent one. What we do host is the rest of the stack: the language models that write and rewrite scripts, and production speech models such as Gemini 3.1 Flash TTS Preview and GPT-4o mini TTS for teams that want a managed voice rather than a self-hosted one.

The useful part is the seam between those layers. If you self-host Ming, your voice model has no availability guarantee, no autoscaling and no second region; the routing layer can make the text half of the pipeline redundant and fail over automatically when a provider degrades. One API key covers 200-plus models at 0% markup on provider list price, and the routing DSL lets you compose a primary path and a fallback into a single call rather than scattering retry logic through your application. Model fusion is the other option worth knowing about: for something like script adaptation, a panel of models answering together often beats one model when the job is judgment rather than generation. None of this makes Ming faster or Sesame licensable. It makes the pipeline around whichever voice you choose survive the voice being the weak link.

The decision, stated plainly

If you want a conversational voice agent in English this week, Sesame Preview is the finished product and nothing in the open-weight world matches its latency or its out-of-the-box personality. Accept that you cannot license the voice, that you are building on a closed app, and that the 4.7 MOS is the ceiling of what you can verify.

If you want a voice inside your own product — cloned, designed, controlled, in seven languages, with MIT terms and no per-character meter — Ming-3.0-tts is the choice, with two conditions attached. Test your own vocabulary before you trust the 93.2% pronunciation score, because the standalone-terms result is much weaker. And plan for the engineering: the inference path is research-grade, there is no streaming story yet, and the English voice catalogue is five voices deep.

The one thing you should not do is read this as a quality contest. Sesame's demo voice probably does sound better than Ming's stock English voices; that is what a curated four-persona product optimized for conversation should achieve. The question is whether you are buying the voice or renting the experience of hearing it. Only one of those is on the menu.