Nuplex 1.0 and Emma: A Video Persona That Answers on a Human's Beat

The most disarming sentence in Nuance Labs' launch post is the one where the company disowns its own benchmark. After presenting a chart in which its new model matches a human's reply timing 89% of the time against 35% for the next-closest system, the post concedes that "benchmarks in the AV full-duplex space are not accurate to the user experience yet." That is not a hedge bolted to the bottom of a marketing page. It is the honest frame for what Nuplex 1.0 actually is: the first public release from a Seattle lab that has spent 18 months building an audiovisual model which sees, hears, reasons and speaks in the same instant, shipped on November 9, 2026 inside a single consumer persona called Emma. What follows separates what the vendor measured, what it says its own measurements cannot tell you, and what a developer can reach today.

The shape of it, in one paragraph. Nuplex 1.0 is a full-duplex audiovisual foundation model — "Nuplex" is the company's contraction of Nuance Full-duplex, a system that listens and speaks simultaneously rather than taking turns. Emma is the first persona built on it: a video character you talk to in a browser with a microphone and an optional camera, whose face and voice are licensed from a real actress, and who is designed to nod, smile, look away and interject while you are still speaking. The consumer experience needs no account. The Nuplex Beta API, which is how anyone other than a curious visitor gets to touch the model, sits behind a sign-up at the company's platform site.

What shipped, and what is still closed

The primary artifact is a research post, "Meet Emma, the First Nuplex Persona," dated November 9, 2026 and carrying a citable BibTeX entry — a researcher's release, not a product keynote. It links two doors: emma.nuancelabs.ai for the persona itself, and platform.nuancelabs.ai for the beta API with "a few different personas." Beyond the default character, the API brief lets you write a persona prompt, follow the conversation as it happens, and let the persona hand longer tasks to your own tools in the background while it keeps talking to the user.

The same evening, a founding researcher posted the persona to Hacker News under the title "Emma, a real-time AI video persona from Nuance Labs" and answered questions in the thread. Three details there are worth having because they are not on the marketing page: that the beta API is the intended developer path, that the team frames the work as betting on naturalness over existing benchmarks, and that early callers were warned Emma "may prematurely end the call" — a safety behaviour rather than a bug, which the research post later explains in detail.

What is not open is most of the commercial surface. The platform subdomain sits behind an authenticated sign-up shell, and no public price list, rate card or model card carrying a parameter count was reachable. So the honest statement about cost is this: talking to Emma needs no account and no payment, and the beta API's terms are arranged at sign-up. Anyone quoting a per-token price for Nuplex 1.0 today is inventing it.

The 89% claim, read the way the vendor reads it

Reply latency is the number every full-duplex system gets judged on, and Nuance Labs argues it settles very little by itself, because latency is a product decision rather than a score. Its own framing splits the field in two. Open models like PersonaPlex and Moshi are tuned to answer as fast as possible and max out the existing benchmarks. Product systems like GPT-Live and G​emini Live take longer, "most likely because they would rather not cut you off." Nuplex 1.0 is placed deliberately between the two.

So the company measured something else: not how fast a system answers, but whether it answers when a person would. On 1,785 questions drawn from its own recordings of real conversations, Nuplex's reply timing overlapped the original speaker's 89% of the time. The next-closest system managed 35%. Every part of that sentence is vendor-reported and reproduced by nobody outside the company, and two of the comparison figures carry a caveat the vendor volunteers itself: GPT-Live and G​emini Live were reached through their public APIs, so their timings include network time that Nuplex's do not.

Take the claim for what it is worth and it is still a specific, checkable thing to have published — a named question set, a named metric, a stated comparison method. The more interesting question is why a lab would lead with a number and then undercut it in the next paragraph. Because the metric measures the thing the product actually sells. A face-to-face conversation is not a fast conversation; it is one where the other party's timing signals attention. Emma's visual cues — a nod while you talk, a gaze shift that claims the turn — are part of how that timing gets produced, and that is the concrete difference between this and a speech-to-speech model with an avatar rendered on top of it.

A face from one actress, and a personality trained to disagree

The least launch-like part of the post is the part about why generated faces look wrong. Ask professional actors to portray the same emotion and you get wide variation, the company notes, citing Le Mau et al. 2021; AI-generated faces instead produce a prototypical expression that is easier to label and more intense than a real posed photograph, citing Kramer et al. 2026. Average many people and you get something muted. Optimise for a label and you get a cartoon.

Nuance Labs' answer was to stop averaging. Emma's face and voice come from one actress who collaborated on the capture, and her mannerisms are modelled on a specific character rather than a generalisation of training data — the long blink for emphasis, the squint, eyes that wander and come back, the small jump when she cannot contain a reaction. The post sets her against prototype emoji-style expressions to make the point visually, and it is the kind of difference you notice in two seconds and cannot un-notice afterwards.

The personality work is the stranger claim. Emma is designed to be curious, quick-witted, to hold opinions and to disagree with you, while being openly an AI. The post is blunt about why that is hard: preference data rewards answers that agree with the person asking, so most assistants repeat what you said, praise it, draw a tidy moral and end on a question. The citation is Sharma et al. 2024 on sycophancy, and the design constraint that follows is that there is only one Emma — one face, one voice, one temperament — not a character you can re-prompt into agreement.

What the beta will refuse to do

Nuance Labs opens its safety section by quoting its first testers, who said two things "often in the same breath: it was amazing, and it was a little unsettling." The restrictions that follow are unusually specific for a launch, and they are the reason several obvious products cannot be built on Nuplex 1.0 this month.

• Disclosure is structural — Emma never claims to be human, developer prompts that try to make a persona claim it are refused, and video served through the API carries a visible watermark.

• No cloning — the beta cannot clone real people, you cannot upload a face or a voice, and a custom face built for an organisation can only be used by that organisation.

• No sexual content, and no AI girlfriends or boyfriends.

• Healthcare and anything involving under-18s, education included, are out of the beta entirely; those enquiries are routed to an enterprise address instead.

• Every developer prompt is screened before a conversation starts. During a call, automated checks end it on expressed intent toward serious violence, slurs or hate speech, sexual content, or weapons and gore on camera — and also if the persona itself asks for a passcode, card details, money or remote access to a device.

• Someone who appears to be in crisis is pointed to help, such as the Suicide & Crisis Lifeline, rather than hung up on. Repeated violations pause new conversations. The whole product is 18+.

The post adds the sentence most vendors leave out: these checks will miss things, and they will sometimes end calls that were fine. That is the right thing to tell people, and it is also an accurate description of where the product is — a carefully fenced beta rather than a platform.

What Emma cannot do yet

The limitations list is long enough to be useful. Emma's reasoning comes from a language model, so she makes confident mistakes and can still miscount the letters in "strawberry." Turn-taking fails where turn-taking is hardest: she can jump in mid-thought, wait a beat too long, or miss a cue a person would catch, and long pauses mid-sentence are the worst case. She is English-only for now. Background voices, a television, speakerphone or a weak microphone all degrade her timing. Streaming visual understanding is early — with the camera on she can discuss what she sees but will miss details and misread scenes. Lip sync, gaze and head motion occasionally slip, especially on a poor connection, and her expression can drift back to neutral mid-conversation. She cannot sing. The company says so with a smiley face, which is a fair signal about how seriously to take the rest of the list.

Who is behind Nuplex 1.0

Nuance Labs is small and explicitly research-first, based in Seattle, Washington. Its three named founders are Fangchang Ma (cofounder and CEO, PhD in robotics and machine learning from MIT, previously an engineering manager at Apple), Edward Zhang (cofounder and CTO, PhD in computer graphics from the University of Washington, previously a senior research scientist at Apple) and Karren Yang (cofounder and chief scientist, PhD in audio-visual synthesis from MIT, also previously at Apple). The company's own framing of the model is a single system that "sees, hears, reasons, speaks, and expresses in the same moment," built by predicting the next audio and visual token rather than the next word — the same autoregressive machinery that learned language, pointed at human behaviour.

Its home page also states the design target plainly: "500ms is the hard floor for conversation latency," alongside four pillars — expressivity and naturalness, real-time interactivity, an engaging personality, and active listening through interjections, nodding and backchannelling. None of those four is a benchmark, and the 500ms figure is a stated floor rather than a measured result, so read it as the constraint the team designs against rather than as a performance claim about Nuplex 1.0.

Where you can actually use it

Availability is geographic and narrower than "global." Emma and the Nuplex Beta API are offered in the United States except Colorado, Connecticut, Illinois and Texas; in the EU; in the UK; and in Canada except Quebec. Talking to Emma requires no account — a browser, a microphone and optionally a camera are the whole setup.

For builders, the interesting architectural fact is that Nuplex sits in front of a language model rather than replacing one. Nuance Labs says Nuplex works with any language model but that the beta serves one of its own, so the reasoning layer under Emma is not something a third party supplies today. What the beta does expose is the seam: the persona can hand longer tasks to your tools while it keeps talking, and those tools can call any model or service you like.

That seam is where a router earns its place, and it is worth being exact about what we do and do not do. OrcaRouter does not serve Nuplex 1.0 — the model is not on our catalogue, and the only way to reach it is through Nuance Labs' own beta platform. What we do serve is the other side of the call: 215 models behind one OpenAI-compatible key at provider list price, passed through with 0% markup, so the model a persona's tools reach for can be swapped by changing a string rather than signing a second contract. Automatic failover covers the case that matters most in a live conversation, where a provider stalling mid-call is worse than a slightly weaker answer, and the routing DSL lets you compose several models behind a single endpoint when a background task wants one model to draft and another to check.

What to watch next

Three things would move this from an impressive demo to a platform. The first is independent measurement: no outside party has run the 1,785-question timing test, and the AV full-duplex space has no shared arena, which is precisely why the vendor could publish its own question set without contradiction. The second is the language list — English-only is the single biggest limit on the beta, and a multilingual Nuplex changes who can build on it. The third is the language model underneath: the moment Nuplex is offered with a bring-your-own LM, the routing question stops being about tools and becomes about the conversation itself. Until then, the fair summary is that Nuance Labs has shipped the most convincing argument yet that conversation is a visual act — and has fenced it carefully enough that nobody can quietly build the wrong thing with it.