What Is Sesame AI? Maya, Miles & the CSM Model (2026)

Sesame is the company behind Maya and Miles — the strikingly natural AI voices that went viral in early 2025 — and the open-weight CSM speech model that developers can download. And yes, the demo people keep passing around is the real thing: it's a free research preview of Sesame's app, not a knockoff. The confusion is understandable, because Sesame is two different products wearing one name — a consumer voice-companion app and an open-source model — and almost every write-up blurs them together.

Here's the part worth knowing before you go hunting: Sesame's Maya and Miles drew more than 1 million users who generated over 5 million minutes of conversation within weeks of the February 2025 demo, per TechCrunch, and the company has since raised a $250M Series B (led by Spark Capital and Sequoia), bringing total funding to about $307.6 million. This isn't a weekend hobby project. It's a well-funded bet from the people who built Oculus.

Maya, Miles, Simone & Charlie: the voice companions, not a TTS tool

Sesame was founded by Brendan Iribe (former Oculus CEO) and Ankit Kumar (former CTO of AR startup Ubiquity6). It emerged from stealth in February 2025 with two demo voices — Maya and Miles — and the thing people reacted to was how conversational they felt: pauses, breaths, filler words, the sense of talking to someone who was actually listening. The iOS app now ships four agents — Maya, Miles, Simone, and Charlie — each with its own voice, personality, and persistent memory across conversations.

This is the single most important distinction, and it's the one the SERP gets wrong: Sesame is a conversational voice AI, not a text-to-speech tool. You don't paste a script and download an MP3. You talk to it, and it talks back in real time. If what you actually want is "turn this document into narration I can save," Sesame isn't that product — a studio TTS service or a read-aloud app is (more on that below). Sesame is closer to a voice-first chatbot you'd keep a running relationship with, which is why the company's longer-term plan is to put these agents into lightweight smart glasses, with hardware expected around 2027.

Is Sesame AI free? How to actually try it

As of November 2026, yes — Sesame is free to try. It runs as a research preview: there's a browser demo at app.sesame.com and a mobile app that, per sesame.com, is available on iOS and Android (the iOS public preview opened across 39 countries in May 2026), with eyewear "coming soon." There is no published pricing or paid tier on the site at the time of writing. That fits the stage Sesame is at: it's gathering conversation data and building toward a hardware product, not monetizing a subscription yet. Treat "free" as "free for now, while it's a preview" rather than a permanent promise.

One caveat worth flagging: because the companions have memory and the whole point is natural back-and-forth, you are handing a startup a lot of voice conversation. That's not unique to Sesame, but it's a reasonable thing to weigh before you spend hours talking to Maya about your life.

CSM: the open-weight model developers can self-host

The other half of Sesame is CSM — the Conversational Speech Model. Sesame released a 1B-parameter variant, CSM-1B, under the Apache-2.0 license on March 13, 2025 (weights on Hugging Face, code at github.com/SesameAILabs/csm). Apache-2.0 is the headline developers should care about: it permits commercial use. That's a materially different license story from a model like Microsoft's VibeVoice, whose card restricts it to research only.

Two things to set expectations straight, both from the official repo. First, CSM-1B is a base generation model — Sesame's own README says it "has not been fine-tuned on any specific voice," so it's a foundation to build on, not the polished Maya voice you hear in the app. The app's companions run on Sesame's own hosted system; CSM-1B is the separate piece you can actually download and fine-tune yourself. Second, it's effectively English-first: the repo warns it has only "some capacity for non-English languages due to data contamination" and "likely won't do well" outside English. Architecturally it uses a Llama backbone with a smaller audio decoder that produces Mimi audio codes — standard, reproducible, the kind of thing a team can fine-tune.

If you're a developer weighing CSM against the rest of the open-weight field, we track it alongside Kokoro, VibeVoice, Qwen3-TTS and the others in the open-source TTS roundup — license, size, languages, and voice cloning compared side by side.

What makes it sound different

The press framing ("the world's most realistic AI voice") is marketing, and "most realistic" is a subjective claim no blind test settles cleanly. The substance underneath it is real, though: most TTS models generate a sentence at a time with no idea what came before. Sesame's pitch is that CSM models conversational context — it takes both text and prior audio as input, so the delivery of the next line is shaped by the tone and rhythm of the exchange so far. That's why Maya's pauses and intonation feel like a reply rather than a read-out. It's a design choice aimed at dialogue, not a benchmark trophy, and it's exactly why Sesame is the wrong tool for static narration and the right one for a live assistant.

Where Sesame fits vs ElevenLabs, Hume, and conventional TTS

Three categories get mashed together in these conversations, so it helps to separate them:

• Conversational voice (Sesame): a two-way companion you talk with. Great for a live assistant; not for producing audio files.
• Studio TTS (ElevenLabs, Amazon Polly, OpenAI): paste text, get a high-quality file. The right pick for audiobooks, videos, and anything you need to save and reuse. See our best text-to-speech comparison.
• Emotional / expressive TTS (Hume): studio-style generation tuned for emotional range. Our Hume AI review covers where that sits.

If you're a business evaluating a hosted conversational layer — a phone agent, a support bot, a voice assistant in your own product — Sesame's app isn't sold as that today, and CSM-1B is a base model you'd have to engineer around. Dedicated agent platforms are the practical answer; our best TTS for voice agents guide compares the options (including ElevenLabs Agents) on latency and real cost per minute.

Who should try it — and who should look elsewhere

Try Sesame if you're curious what a genuinely natural voice companion feels like, you want to experiment with a free preview, or you're a developer who wants an Apache-2.0 speech model to fine-tune. Look elsewhere if you need narration you can download (use a studio TTS), if you need reliable multilingual output (CSM is English-first), or if you're shipping a production voice agent this quarter and can't wait on a preview-stage consumer app. The honest one-line test: do you want to talk to a voice, or do you want a voice to read to you? Sesame is built for the first question. Most people searching "AI voice" actually want the second.