Phonon 31M

Gradium's 31M English speech model in a 64.5 MB file. Apache 2.0, streams audio, runs on phones, browsers and plain CPUs.

In short. Phonon 31M turns English text into natural speech and fits in a 64.5 MB file, small enough for a phone, a browser tab or a laptop processor. The Apache 2.0 license allows commercial use. Gradium's larger multilingual Phonon models are not free for commercial use.

Summary

Phonon is a family of on-device text-to-speech models from Gradium, a Paris voice-AI company spun out of the Kyutai lab. The family was announced on 21 October 2026. Phonon 31M is the one released under Apache 2.0: the weights, tokenizer and voices carry that license, and the runtime is a separate project licensed MIT or Apache 2.0. It speaks English only and outputs 24 kHz mono audio. Inside, a flow-matching transformer predicts audio features one 80-millisecond frame at a time and the decoder of Kyutai's Mimi codec turns them into sound, so it streams audio while it is still generating. It builds on Kyutai's Pocket TTS. The model has 31 million parameters, not counting the audio tokenizer, and the weights file, model.q8.gguf, is 8-bit quantized and 64.5 MB.

The runtime covers more platforms than most small speech models. Gradium publishes it for Python (pip install ptts), Rust, C and C++, Swift and Core ML for Apple devices, Android, and the browser through WebAssembly (phonon-tts on npm). It also ships an OpenAI-compatible server, a WebSocket server and a Home Assistant setup. The English model comes with 37 built-in voices, and Gradium says it starts speaking in about 7 ms on a laptop CPU. Cloning a new voice or designing one from a description is not part of the open weights and goes through Gradium's API and studio.

On Gradium's own English test, with 468 raters and 8,290 blind pairwise votes, Phonon 31M scored an Elo of 2,098, against 2,043 for Kokoro, 2,034 for NeuTTS and 1,939 for Magpie. On a hard set of sentences full of numbers, dates and codes, its word error rate was 1.5%, against 3.5% for Kokoro, 8.7% for NeuTTS and 28% for Magpie. On the Seed-TTS sentences its 0.51% trailed Kokoro and the larger Phonon 40M at 0.45%. These are the developer's own results and we have not reproduced them.

The license split matters. Only the English 31M model is Apache 2.0. The Phonon 40M models, one each for English, French, German, Spanish and Portuguese plus a single multilingual model, are CC BY-NC 4.0, and Gradium points commercial users to its sales team. If you need a language other than English under a permissive license, Phonon does not cover it yet.

Good for

  • Private, offline speech on phones, laptops, single-board computers and web pages
  • Accessibility tools, games and assistants that cannot wait for a cloud round trip
  • Commercial English products, since the weights, tokenizer and voices are Apache 2.0
  • A very small download, at 64.5 MB

Watch out for

  • English only, and text in other languages is read with the wrong pronunciation
  • Voice cloning and voice design are not in the open weights
  • The multilingual Phonon 40M models are CC BY-NC 4.0, so they are not free for commercial use
  • Very long input is split into sentences, so intonation does not carry across the whole text
  • Choose the language setting to match the text, because number and date handling depends on it
  • The quality and speed figures are Gradium's own

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Phonon 31M. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
phonon-31m-en (Q8 GGUF) INT8 31M 0.065 GB 0.0747 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Phonon 31M?
Phonon 31M turns English text into natural speech and fits in a 64.5 MB file, small enough for a phone, a browser tab or a laptop processor. The Apache 2.0 license allows commercial use. Gradium's larger multilingual Phonon models are not free for commercial use.
Can I use Phonon 31M commercially?
Yes. Phonon 31M is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Phonon 31M open source?
Phonon 31M is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Phonon 31M on my own computer?
Yes. The smallest variant, phonon-31m-en (Q8 GGUF), needs about 0.0747 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Phonon 31M released?
Phonon 31M was released by Gradium on October 21, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Phonon 40M Gradium 40M – CC BY-NC (Non-commercial) 0.0858 GB Laptop or phone
Pocket TTS Kyutai 18M – 100M – CC BY 4.0 (Permissive) 0.127 GB Laptop or phone
Kokoro hexgrad 82M – Apache 2.0 (Permissive) 0.38 GB Laptop or phone
Supertonic 3 Supertone 99M – OpenRAIL-M (Conditional) 0.46 GB Laptop or phone