Sesame Preview vs NVIDIA NemotronLabs VoiceChat 11B: The Voice You Can't License, and the Voice You Can't Ship

Two of the most talked-about voice models of 2026 have something in common that none of the launch coverage will tell you: you cannot put either one in a product. Sesame Preview is the free consumer assistant whose conversational voice made TestingCatalog call it "one of the best voice modes available on the market," and the production model behind it has no API, no model ID, and no license you can buy — the company has simply not released it. NVIDIA NemotronLabs VoiceChat 11B is the mirror image: the weights are public, the full-duplex design is a genuine first, and the license says research purposes only, so nobody can legally ship it either. One is a voice you can hear but not rent. The other is a voice you can hold but not sell. This is a comparison of two locked doors, and the useful question is which lock you are standing in front of.

Two locked doors, with different locks

Start with the shape of each, because the names hide how different the two products are. Sesame Preview is an app, not an API. It ships four agents with distinct personalities — Maya (warm and creative), Miles (laid-back and sharp), Simone (curious and intellectual), Charlie (witty and warm) — each with its own voice and its own memory that syncs across web and mobile when you are signed in. It searches the web, goes on deep dives, takes notes and reminders, and sends a summary after each call. The preview is free with no paid tier, English-only, capped at 30 minutes per call when signed in (five otherwise), and it deliberately does not execute tasks: it will brainstorm and research, but it will not book your flight. There is no API key anywhere in the product — the production voice is an app feature, not an endpoint.

NVIDIA NemotronLabs VoiceChat 11B is the opposite shape: a checkpoint, not a product. It arrived on Hugging Face on August 3, 2026 with no announcement — no launch post, no tech report, no tweet; the model card is the announcement. It is an 11B-parameter full-duplex speech model (we parsed the safetensors header and counted 11.095 billion parameters across 1,626 tensors) built as one stack: a Fast Conformer encoder chewing 16 kHz audio in 80 ms frames with zero right context, a Nemotron Nano 9B v2 backbone doing the reasoning, an NVIDIA T​TS decoder writing 22.05 kHz audio, an RNN-T decoder transcribing the user, and a separate output channel that emits tool-calling scripts without contaminating the spoken reply. It listens and speaks at the same time, can be interrupted mid-word, and NVIDIA bills it as the first open full-duplex model with tool calling. Whether that billing survives contact with reality is the subject of its own section below.

The benchmark where they diverge — two candidates for "best conversation," two broken evidence standards

Both models stake their entire reputation on the same thing: conversational quality — turn-taking, interruption, natural timing, the feel of a real exchange. That is why they belong in one headline at all. It is also where the comparison goes strange, because neither candidate can be scored properly.

Sesame Preview has no number anywhere. The production model is unreleased and unlisted — no model ID, no entry on Artificial Analysis' speech-to-speech leaderboard, no latency target, no benchmark of any kind. TestingCatalog's "one of the best voice modes available on the market" is reviewer consensus, not a measurement, and there is no way to blind-A/B it against anything. The only Sesame model anyone can independently run is the open-weight CSM-1B base, and that is a text-to-speech checkpoint, not the app's voice.

NVIDIA NemotronLabs VoiceChat 11B has the opposite problem: it has numbers, but the numbers are split across two evidence standards that do not agree. NVIDIA reports 448 ms to take a smooth turn, 480 ms to yield when interrupted, and a perfect 1.0 takeover rate on user interruption — all vendor-measured on NVIDIA's own Full-Duplex-Bench 1.0, none independently reproduced. The independent measurements of this family are filed under a different name and parameter count — "Nemotron 3 VoiceChat (V1)" at 12B, a label NVIDIA has never reconciled with the 11B checkpoint on Hugging Face — and they are unflattering on the understanding side: 29.2% on Big Bench Audio per Artificial Analysis, against 37.0% on NVIDIA's own card. On VoiceBench the family scores 58.10, the best open full-duplex result on the board. So the same model is simultaneously a leader among open full-duplex checkpoints and roughly three-to-one behind hosted realtime leaders at understanding what it hears.

The divergence, then, is not which is better. It is that the two candidates for the best conversation of 2026 are being judged on opposite and equally incomplete evidence. The voice reviewers say feels the most human has no measurement at all, and the voice with actual leaderboard entries is the one whose weaknesses are public. You cannot compare a reputation to a number, and here there is only one number — and it is not the flattering one.

Access — an app-store download, or an 80 GB build

If you want to try either one, the starting line differs in a way that matters more than any spec.

Sesame Preview is a download. The official page reads "Preview available now on iOS and Android," and the Android build has been rolling out since mid-July (the Android beta added image upload and voice memos). There is no API key anywhere. A developer who wants Sesame's actual voice has nothing to call — the only Sesame model reachable through an API is the open-weight CSM-1B base, listed at $7 per million characters, and that is the architecture behind the app, not the app's voice: a text-to-speech checkpoint with a 4K context window and no conversational abilities of its own.

NVIDIA NemotronLabs VoiceChat 11B is not a download you can simply run either — it is a build you have to stand up. Hugging Face states the model "isn't deployed by any Inference Provider"; NVIDIA has not hosted it; and the config.json in the repo is a NeMo training config, so there is no Transformers checkpoint you can point AutoModel at. The supported path is a conda environment pinned to Python 3.12, PyTorch 2.10 and Transformers 4.56, with causal-conv1d and mamba-ssm compiled from source, on an 80 GB GPU (A100, H100, H200, B200, or RTX 6000) running vLLM — or NVIDIA's NGC NIM container, which exposes an O​penAI Realtime-compatible WebSocket at /v1/realtime once you are on that hardware. The download counter tells the access story: roughly 80 downloads in the model's first month against 107 likes. People bookmarked it; almost nobody ran it. One honest note for the researcher who does: the reasoning backbone this checkpoint builds on, Nemotron Nano 9B v2, is itself a hosted model you can call today — the full-duplex model around it is not, and that gap is the whole point.

Price — a free app, free weights, and the 80 GB bill

Both models are "free," and the word does the same amount of misleading work in both cases.

Sesame Preview is genuinely free — no paid tier, no ads, no data selling — because it is a preview you are being invited to test, and Sesame's monetization is a later problem. NVIDIA NemotronLabs VoiceChat 11B's weights cost nothing to download, but the hardware does: a continuously-running H100-class instance at roughly $2–3 an hour lands near $1,500–2,200 a month before you have written any application code, hired anyone to keep it up, or served a second concurrent conversation — and because the license forbids commercial use, that is money you can only spend on research. Between them sits CSM-1B at $7 per million characters, a hosted price for a model whose default is self-hosting.

The honest yardstick for the day either of these becomes buyable: whatever hosted voice model you land on, the provider's list price is what you should pay, with no routing markup on top — the pass-through that makes a vendor price cut show up in your bill the same day rather than after a contract cycle. Neither model in this piece is callable through OrcaRouter: Sesame's production voice has no API at all, and NVIDIA NemotronLabs VoiceChat 11B is not callable through anything. The yardstick is for the voice you can actually buy, and it is exactly the standard that makes a $7 per million-character T​TS base or a future Sesame-quality API easy to price honestly.

Memory — the app that remembers vs the model that forgets

Here the gap is a canyon, and it is the most concrete reason the two are not substitutes. Sesame Preview the app remembers: each agent carries per-agent memory that syncs across web and mobile when you are signed in, calls can run to 30 minutes, and Incognito Mode reads past memories without creating new ones. The open CSM-1B, by contrast, has a 4K context window — roughly three to four minutes of text — and no ears to hear you anyway.

NVIDIA NemotronLabs VoiceChat 11B holds at most about two minutes of audio context — the training dataloader caps utterances at 142.39 seconds, and the card warns that conversation beyond a two-minute window may not be retained. For a support call that needs to remember what was agreed four minutes ago, that ceiling is disqualifying on its own. Add a single fixed voice (Aria) with no cloning in the released checkpoint, and the model that is open about its architecture is also the model that cannot remember a customer or change its own voice.

What each one is genuinely bad at

Collected, because a comparison that stops at strengths is marketing:

• Sesame Preview: no API — you cannot put this voice in a product, and the production model is unreleased and unscored. English-only, no task execution, a 30-minute call cap, and no independent benchmark of any kind. The one open model, CSM-1B, has a 4K context window, no tool-calling, no listening side, and is not the app's voice.

• NVIDIA NemotronLabs VoiceChat 11B: a research-only license (OpenMDW v1.1) — you can download, study, and fine-tune it, and you cannot ship it, and neither can anyone who builds on it. No hosted endpoint, so "trying it" means an 80 GB GPU and a from-source build. English-only, a ~2-minute memory ceiling, one fixed voice. NVIDIA states it may underperform its own backbone on knowledge and instruction-following, with multi-step reasoning and arithmetic called out as weak. It can cut you off mid-sentence, degrade into non-recoverable gibberish or self-talk after several turns, drop words in transcription, and it is explicitly unsuited to noisy or reverberant rooms. Tool calling works only with ASCII-only prompts and responses, tops out at five tools per session, and its argument accuracy (44.2%) and first-try pass rate (33%) mean it picks the right tool more reliably than it fills the tool in.

Who should pick which

Because neither is callable, the honest answer starts from what you are trying to do.

• Experience the best-reviewed voice of 2026: Sesame Preview, free, today — as an app. The four agents, the interruption handling, the drive-mode usability reviewers keep citing: that is a consumer product, and its value is precisely the thing you cannot measure.

• Study full-duplex speech modeling: NVIDIA NemotronLabs VoiceChat 11B is the more interesting download in its category — an 11B open checkpoint with a working tool-calling channel, roughly 550,000 hours of blended training audio, and the best open full-duplex VoiceBench result yet, all at a cost of zero dollars for the weights. The research-only license is not a constraint on that work.

• Build on Sesame's architecture: CSM-1B is the only Sesame model a developer can actually call — $7 per million characters, Apache-2.0, zero-shot voice cloning — with the honest caveat that it is a base text-to-speech model, not the production voice the app uses.

• Ship a voice product this quarter: neither of these. You need a hosted realtime speech-to-speech model with a commercial license, and the safe way to adopt one you have not bet a production path on is to route to it beside a proven model with automatic failover, so an unproven voice never takes down a live call. One API across 200+ hosted models at provider list price, with no markup, is what makes trying a newly-available voice a configuration change rather than a rewrite.

The pattern is worth naming because it will recur. Both of these models are one event away from becoming callable — Sesame the day it ships a model ID or API, NVIDIA NemotronLabs VoiceChat 11B the day NVIDIA publishes a commercial license or anyone hosts it. When that happens, the right move is not to rip out your current voice stack. It is to add the new voice beside the old one and route with failover, exactly as you would for any new, unproven model. These are the two best unshippable voices of 2026; the infrastructure to ship the third one already exists.

What would change this comparison

Four events would rewrite this piece. An API or model ID for Sesame's production voice — the moment that happens, Sesame stops being an app and becomes the entrant the hosted voice-API market has been waiting for. A commercial license or a hosted endpoint for NVIDIA NemotronLabs VoiceChat 11B, which would turn a research artifact into a real product option overnight. An independent score for the Sesame voice — a listing on Artificial Analysis' speech-to-speech board would give the "best voice" claim something to be checked against. And a tech report from NVIDIA reconciling the 11B checkpoint with the 12B "Nemotron 3 VoiceChat (V1)" entries the leaderboards measure, which is the precondition for trusting any of the family's numbers. Until any of that lands, the accurate summary is simple: the two most interesting conversational voices of 2026 are both locked — one by a company that will not sell, the other by a license that will not ship.

FAQ

Can I use either model's voice inside my own app?

No, and the two reasons are the shape of this comparison. Sesame Preview's production voice has no API, no model ID, and no endpoint — it exists only as the app. NVIDIA NemotronLabs VoiceChat 11B is open-weight but research-only (OpenMDW v1.1), self-hosted on an 80 GB GPU, with no hosted inference provider. The only Sesame model a developer can call is CSM-1B at $7 per million characters, and it is a text-to-speech base, not the app's voice.

Which of the two actually sounds more human?

Unknown, on different grounds for each. Sesame Preview has no independent score at all — TestingCatalog's "one of the best voice modes available on the market" is reviewer consensus, not a measurement. NVIDIA NemotronLabs VoiceChat 11B's conversational timing (448 ms turn-taking, 480 ms barge-in yield) is NVIDIA-reported and unreproduced, and its independent Big Bench Audio figure (29.2%, per Artificial Analysis, filed under "Nemotron 3 VoiceChat (V1)") measures understanding, not how pleasant the voice is. One has a reputation you can hear; the other has numbers that answer a different question.

Is NVIDIA NemotronLabs VoiceChat 11B really free?

The weights are free; the model is not cheap. Running it means an 80 GB GPU (roughly $2–3 an hour on demand, $1,500–2,200 a month if kept continuously warm) and a from-source build, and the OpenMDW v1.1 license restricts it to research purposes only — so even after that spend, you cannot legally put it in front of customers.