Ming-3.0-audio-flash vs Sesame Preview: One Hears Everything, the Other Only Needs to Sound Human

These two are not rivals. Ming-3.0-audio-flash is a 126-billion-parameter audio-language mixture-of-experts under MIT whose five stated jobs are speech recognition, long-audio understanding, perception of non-speech sound and music, audio reasoning, and — as an optional fifth — speaking back. Sesame Preview is a consumer voice assistant with four personalities, an iOS and Android app, and no API at all. The second one does not need to know what a door slamming sounds like. The first one was built for exactly that. Putting them side by side is useful only because it forces the question most audio-model comparisons skip: are you buying an ear, or a mouth? 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 timing makes the mismatch sharper. Ming-3.0-audio-flash landed on Hugging Face on 3 September 2026 without an announcement and sat there for six weeks before the serving code, examples and rewritten card went up between 15 and 19 October — the GitHub repository was created on the 15th, and the card was still being edited on the 19th. The listing shows 22 downloads. Sesame Preview has been shipping in public since the spring, and its downloadable research artifact, CSM-1B, has been on Hugging Face since March 2025 with roughly 130,000 downloads in the last month and about 2,500 likes. One has a user base and no repository. The other has a repository and no users yet.

What each one is made of

Ming-3.0-audio-flash combines an audio encoder and projector with a mixture-of-experts language backbone, and an optional Talker module that converts the model's text output into speech. Total parameters are 126 billion, with 6.5 billion activated per inference step — the Hugging Face sidebar lists 128B, a discrepancy the card does not reconcile. Inputs are audio plus a text instruction. Outputs are text, or text with synthesized speech. It handles Chinese and English, with accents and dialects called out explicitly.

Sesame Preview is not described at that level of detail anywhere public. What is public is the app: a voice interface with four named personalities, memory, connectors, and since October 2026 the ability to orchestrate other agents. What is also public is CSM-1B, Sesame's Apache-2.0 research checkpoint — a Llama backbone predicting the first audio codebook and a smaller depth decoder predicting the rest, decoded through Kyutai's Mimi codec at 24 kHz, English-only. The card calls it "a base generation model," and Sesame has been explicit that a fine-tuned derivative of it powers the interactive demo rather than the checkpoint being the demo.

• What you can download — Ming: the full 126B checkpoint, processor, tokenizer, Talker assets. Sesame: CSM-1B, a 1B-parameter base model that is not the voice in the app.

• What you can call — Ming: an OpenAI-compatible endpoint on vLLM-Omni or SGLang, once you host it. Sesame: an app. No endpoint, no model ID for the production voice.

• What it listens to — Ming: speech, dialects, noise, multiple speakers, non-speech sound, acoustic scenes, music. Sesame: your voice, in a conversation.

• What it says — Ming: text, optionally speech through built-in or reference-cloned voices. Sesame: speech built for breath, hesitation and interruption.

• Languages — Ming: Chinese and English. Sesame: English officially; the app's own guidance lists French, Italian, German, Spanish, Chinese, Japanese and Korean as "early."

• Licence — Ming: MIT. Sesame: app terms of service; Apache-2.0 only on the research checkpoint.

The evaluation asymmetry is total

With Ming-3.0-audio-flash you can generate the same sentence a hundred times and listen to it. You can feed it the same noisy meeting and check the transcript. You can build your own error rate on your own audio, and that error rate is a number you own.

With Sesame Preview you cannot generate anything. There is no endpoint, no checkpoint for the production voice, no published benchmark for it. You can listen to the app and form an opinion, and that opinion is the entirety of your evidence. The research checkpoint CSM-1B is testable — you can run it, fine-tune it, measure it — but it is explicitly not the model behind the voice that made Sesame famous.

This is the rare matchup where the open model is the only one you can evaluate, even though the closed one is the only one with a real user base. That inversion is worth sitting with, because the usual argument runs the other way: closed products are supposed to be the ones with track records.

Where the two actually touch

There is exactly one overlap, and it is the Talker. Ming-3.0-audio-flash can produce speech, and its voice story is more configurable than a four-voice app: the vLLM-Omni path ships built-in IP voices, the model directory carries preset reference voices with matching transcripts, and you can pass any reference WAV plus its exact transcript to clone. Streaming speech with latency metrics has a working example. The SGLang path does not do speech at all — it is text-only — which is worth knowing before you pick a serving stack.

But configurability is not the same as quality, and the Ming repository publishes no speech-quality evaluation at all. Its three benchmark panels cover speech recognition, audio intelligence and text. There is no MOS score, no preference test, no prosody comparison. So the honest position on "Ming sounds better than Sesame" is that nobody has measured it, and the people best placed to have an opinion — Sesame's users — have never heard Ming.

What Ming's Talker does give you that Sesame cannot is voice ownership. A reference clip and its transcript, and the model speaks in that voice, on your hardware, under MIT, with no third party in the loop. If your product's voice is a brand asset, that is not a feature comparison. It is a different category of thing.

Hardware, and why it decides most of this

Ming-3.0-audio-flash's reference serving profile for the full 128K context with speech output is tensor parallelism across eight 80 GB H800 GPUs. Four GPUs get you a validated SGLang profile, but text only. The repository notes that the full-audio profile gives the Talker its own memory budget on GPU 7.

Sesame Preview has no hardware requirement because it has no deployment. That is not a joke at its expense; it is the actual value proposition. A team that wants an expressive voice in a mobile app and does not want to think about GPUs can ship Sesame-shaped behaviour through a vendor and never see a YAML file.

The cost comparison that follows is not close. Eight H800s is a capital decision with a depreciation schedule. An app is a download. The only reason to compare them is that both can talk, and the cost of talking is not the interesting variable — the cost of hearing is, and Sesame does not sell that.

What each one is genuinely better at

Ming-3.0-audio-flash is better at everything that involves understanding audio rather than producing it. Long recordings, multiple speakers, background noise, music, acoustic scenes, questions about what was said and how it was said. Its own card claims complex recognition across dialects and accents; its benchmark panels are vendor-run and unaudited, but they exist, which is more than the competition in this matchup offers.

Sesame Preview is better at being a product. It has memory, connectors to mail and calendar and drive, its own browsing and acting surface, MCP support over OAuth, and since October the ability to orchestrate other agents. None of that is a speech-model feature; it is an assistant, and speech is the interface. Ming has none of it and does not claim to.

There is also a latency argument that nobody has settled. A conversational assistant lives or dies on turn-taking — Sesame built its reputation on hesitation and interruption timing convincing enough that early listeners assumed a person was on the line. Ming-3.0-audio-flash's interface is request-and-response: audio plus an instruction in, text or speech out. Streaming examples exist, but streaming a response is not the same as handling barge-in, and the repository makes no duplex claim.

If you are actually choosing

You are probably not choosing between these two, and it is worth saying so plainly. The realistic decisions are "Ming-3.0-audio-flash versus a hosted audio-understanding API" and "Sesame Preview versus whatever voice your app already uses." This pairing exists because both are audio models with recent news, not because a buyer would ever weigh them against each other.

That said, the adjacent question — should I self-host an open audio model at all — is the one both products force. Ming makes the case for ownership: MIT, inspectable, modifiable, no vendor in the request path. Sesame makes the case against building: an app that works, with capabilities your team would spend a year reproducing. Most teams should run the hosted version until the hosted version fails them in a way they can name.

OrcaRouter's role in that decision is narrow and worth stating precisely. We do not host Ming-3.0-audio-flash, and we do not host Sesame Preview — one is a checkpoint you run yourself, the other is an app. What a routing layer gives you is the comparison framework: more than 200 models behind one key at provider list price with no markup added, so the hosted audio models you are actually weighing against a self-hosted Ming can be A/B'd on the same traffic, with automatic failover if one path degrades. That is how you find out whether the eight-GPU bet pays, without making the bet first.

The open question

Ming-3.0-audio-flash became runnable four days ago and has 22 downloads. Its card was edited the day this was written. Its benchmark numbers are images, its speech quality is unevaluated, and its serving profile assumes a rack most teams do not own. None of that is disqualifying — it is what week one of a quiet release looks like, and early what-we-know pieces are the right response to it.

Sesame Preview, meanwhile, keeps shipping features into a product with no published model behind them. The gap between what you can hear and what you can inspect is the whole story of that company, and it has not changed in a year.

The thing to watch is not a benchmark. It is whether either side closes its gap: Ming getting an operator and a latency number, or Sesame publishing a checkpoint that is actually the voice in the app. Whichever happens first tells you which of these two approaches is winning.