Kyutai Shrinks Pocket TTS to 1.9M Parameters

On October 20, 2026, Kyutai released a family of English Pocket TTS models that run from 18.2 million parameters down to 1.9 million — small enough to sit on a phone or a microcontroller, and a fraction of the 100-million-parameter model the lab shipped in January. The release arrived under a title that doubles as its thesis: “The tiny TTS was actually 10x too big.”

What Shipped

Version 4.0.0 of the pocket-tts package adds six English checkpoints, selected at run time with the --language flag: english_18m_2026-10, english_7m_2026-10 and english_2m_2026-10, plus a _refl variant of each. Kyutai lists the parameter counts as 18.2M, 6.9M and 1.9M, decoder included, and marks the two smallest — the 1.9M pair — as experimental.

The baseline for comparison is the original Pocket TTS, a 100M-parameter model that Kyutai released in January 2026 to run in real time on a CPU. Its published figures were roughly 6× real time on a MacBook Air M4 using two cores, with about 200 ms to the first audio chunk and voice cloning from a short reference recording. The new models keep the same inference code and command-line interface, with the weights swapped underneath.

The release also ships something the project had lacked: standalone binaries. Prebuilt executables for Linux, macOS and Windows expose the same command-line interface as the Python package and share its Hugging Face cache, so the model can be run without a Python environment at all. They are built on Gradium’s xn-ptts Rust port. The repository code remains MIT-licensed.

How It Got Smaller

Kyutai’s post says the reduction came from three places — the backbone, the sampling head and the decoder — rather than from distillation into an existing architecture. The release notes give one concrete piece: a “drifting” sampler head, trained with a different objective from the earlier models, which ships alongside the existing checkpoints as english_drifting_26-09. Every new option defaults to the old behaviour, so the original models are unchanged.

The engineering headroom was real. A 100M-parameter model that already runs faster than real time on two CPU cores was not constrained by compute; it was constrained by the assumption that a flow-based speech model needs that many parameters to sound natural. Shrinking the sampling head matters disproportionately because it runs once per generated frame, so its cost is paid on every frame of audio rather than once per request.

The Quality Question

Kyutai’s own framing is careful. The largest of the new models, at 18.2M parameters, “matches the quality of previous Pocket TTS releases”, according to the post; the smaller ones “explore the Pareto frontier”. That is a claim about the 18.2M model, not about the 1.9M one, and the experimental label on the smallest pair is the lab telling readers the same thing.

Independent evidence arrived the same day from an unusual direction. Gradium, a voice-AI company that describes Pocket TTS as the starting point for its own on-device work, published a blind listening study covering 1,189 raters and 17,270 votes. In its English results, Pocket TTS at 100M parameters scored 2056 Elo, a 15.6M-parameter Pocket TTS model scored 1997, Kokoro scored 2043 and Gradium’s own 40M Phonon model scored 2117. The study is a competitor’s, and a 59-point gap is inside the range where listener panels disagree, but it does not support the idea that a tenth of the parameters is free.

What This Means

The interesting number here is not the Elo score but the deployment envelope. At 1.9M parameters a speech model stops being a service and becomes a component: it fits in an app bundle, in a browser tab through WebAssembly, or on a device that never has a network connection. Kyutai’s release notes show the project already has that audience — community ports to ONNX, GGUF, MLX, sherpa-onnx and Android LiteRT are listed in the repository, and the new Rust binaries make the CPU path first-class rather than a Python workaround.

The caveat is the same one that follows every size claim. Quality at 1.9M parameters is measured by the lab that trained it, on English, with the decoder counted in the total and the model marked experimental. For a reading app or a game character, that trade is easy to accept. For anything a user will listen to for an hour, the 18.2M checkpoints are the ones with a quality claim attached — and even that claim is Kyutai’s own.

This post was drafted with AI assistance and reviewed by RITS staff.

Sources

• Kyutai: “The tiny TTS was actually 10x too big” (October 20, 2026)
• GitHub: pocket-tts v4.0.0 release notes (October 20, 2026)
• GitHub: kyutai-labs/pocket-tts repository and README
• Hugging Face: kyutai/pocket-tts model card (gated)
• Gradium: “Open-Sourcing Phonon: A Suite of Frontier On-Device TTS Models” (October 20, 2026) — blind listening Elo results