Fish Audio Live 3 Is Production-Ready and Closed: What vLLM-Omni's Offer Actually Asks For

The most revealing sentence about Fish Audio Live 3 was not written by Fish Audio. When the company announced the model on 10 November 2026, the reply that travelled furthest came from the vLLM project's account: congratulations on shipping a full-duplex voice model that listens while it speaks, a wish that it might be open-sourced one day, and an offer that vLLM-Omni would be glad to serve it once it is. Read that reply twice and you have the shape of the whole release, because it contains both halves of the story at once — a genuine capability advance, and an absence that a serving framework notices immediately.

Fish Audio Live 3 is a full-duplex speech model: it keeps listening while it generates speech, and it is sold as production-ready on day one through Fish Audio's Agent Platform, through the company's own API, and as an on-premises deployment for enterprises. What it is not is downloadable. No weights, no licence, no repository, and — as of 11 November — no model identifier in Fish Audio's API documentation, which still lists s2.1-pro, s2.1-pro-free, s2-pro, s1 and three transcribe models and nothing called Live 3. Every performance number below is Fish Audio's own and none has been independently reproduced.

What shipped, and what "production-ready" is doing in that sentence

The technical claim in the announcement is narrower and more interesting than the marketing line. Fish Audio says the model handles conversational timing inside the speech model rather than across a pipeline of separate recognition, reasoning and generation stages, and that it learned when to acknowledge, wait, answer or stop from recorded human conversations rather than from hand-written silence thresholds. A separate reasoning layer holds the agent's instructions and tools. Three sample calls ship with the post — finishing an order, cutting in, and backchannelling during a read-back — and they are the closest thing to evidence in the release.

What "production-ready" means in practice is that the commercial wrapper exists on launch day: a self-serve platform with a $5 credit and no card, a documented agent builder, telephony, and an enterprise on-premises option. It does not mean the model has been through anything a buyer would call acceptance testing. There is no published evaluation of any kind — no word error rate, no task success, no turn-taking score, no latency distribution. That is not unusual for a speech model on its first day, but it is worth saying plainly, because a launch post that leads with milliseconds invites readers to treat a floor as a measurement.

Three numbers, and only two of them come from Fish Audio

• Reaction speed — 186 milliseconds, Fish Audio's own figure, described in the context of short reactions such as an "mm-hmm" or a quick confirmation while the full answer is prepared. It is a floor on an acknowledgement, not a time-to-first-audio for an answer.

• Price — $0.035 per minute on the Agent Platform, promoted as 50% off a $0.07 list rate, with the offer stated to end 31 December 2026. That per-minute rate bundles speech recognition, the reasoning model and the voice; premium LLM tokens bill at their own token rate, phone calls add $0.01 per minute, and phone numbers cost $1.20 a month.

• Independent scores — zero. Artificial Analysis runs a speech-to-speech board that ranks Grok Voice 3.0, GPT-Live-1, Gemini 3.8 Live, Qwen Audio 3.0 Realtime and a dozen others on reasoning, conversational dynamics, task success and time to first audio. Checked on 11 November 2026, that board carries no Fish Audio duplex row at all.

The first two numbers are the ones a buyer can act on. The third is the one a buyer will eventually need, and it is the reason the price deserves a second look rather than a first one: a promotional rate with a published end date is not a cost model, it is a cost estimate with an expiry printed on it.

Why a serving framework cares about a closed model

vLLM-Omni's interest is not sentiment. Serving frameworks exist to run weights you can load, on hardware you control, at a marginal cost that is your electricity bill rather than someone's per-minute meter. A closed model is not a customer for that. The offer in that reply is therefore best read as a request addressed to a future release: publish the checkpoint and there is an inference path waiting for it.

What makes the absence pointed rather than inevitable is Fish Audio's own record. The company has shipped open weights before — Fish Audio Drama 3, its expressive text-to-speech model, is available as a downloadable checkpoint and has been scored on open-weights speech arenas. So the closed distribution here reads as a product decision about the live-agent business, not a house policy about openness. The on-premises option muddies the question further: enterprises can run Live 3 inside their own perimeter, which is not the same as being able to read, fine-tune or re-host it, and Fish Audio has not said which of those the on-prem deal includes.

The English-only line sits next to a "30+ languages" claim

Two sentences on Fish Audio's own properties pull in different directions. The announcement says, in as many words, that Live 3 speaks English today. The Agent Platform page, on the same site, advertises 2,000,000+ voices across 30+ languages and a caller "answered in seconds, day or night, in the caller's language." Both can be true at once — a large multilingual voice library and a speech recognition layer can cover thirty languages while the duplex model underneath speaks one — but only one of those sentences describes the thing you are buying when you buy Live 3. If your callers are not English-speaking, the interesting question is not what the platform can pronounce but when the duplex model catches up.

What you can call today, and through what

Access runs through Fish Audio's own API and the Agent Platform, with the on-premises route reserved for enterprise contracts. There is no third-party serving of the model, which is the direct consequence of the missing checkpoint: you cannot rent Live 3 from an infrastructure provider because there is no artefact to deploy. Anyone describing the model as "available through partner platforms" is describing the platform's voice library, not the duplex model.

The half of an agent stack that is portable is the reasoning half, and on the Agent Platform that half is a separate line on the bill. That is where a router earns its keep rather than as a way to reach the speech model: OrcaRouter carries more than 200 models behind one key at provider list price with zero markup, so the text model doing the thinking can be swapped, pinned or failed over without touching the speech contract, and automatic failover keeps a call alive when the reasoning backend errors or times out. Fish Audio Live 3 is not on that list — there is nothing to host — and no part of this piece should be read as a claim otherwise.

What would change this piece

Four events, in rough order of how much they would move the picture. A checkpoint, which is what the vLLM-Omni reply is asking for and what would put Live 3 on the same shelf as Drama 3. A row on an independent speech-to-speech board, which would turn 186 milliseconds from an assertion into a data point with context. A published model string and rate in Fish Audio's own API documentation, which would make the thing quotable in a procurement document. And 1 January 2027, when the $0.035 promotional rate either survives or does not.

Until then the honest summary is that Live 3 is the most interesting voice model released this week and the least documented one: a real architectural argument about where turn-taking should live, wrapped in a commercial product that has published its price and nothing else. That is a reasonable thing to ship. It is not yet a thing you can evaluate, and the difference between those two sentences is exactly what a serving framework, an independent benchmark and a procurement team each wait for.