Phonon 40M

Gradium's 40M on-device speech models for English, French, German, Spanish and Portuguese, in 75 MB files. CC BY-NC, gated.

In short. Phonon 40M turns text into natural speech in five languages and is small enough for a phone, a browser tab or a laptop processor. There is one model for each language and one multilingual model that covers all five. The weights are free for non-commercial use only, and Hugging Face asks you to accept terms before you can download them.

Summary

Phonon 40M is the larger half of Gradium's Phonon text-to-speech family, released together with the Apache 2.0 Phonon 31M. It comes as one model per language for English, French, German, Spanish and Portuguese, plus a single multilingual model that covers all five with the same weights. Gradium positions it as the one to pick when expressivity and accuracy on hard text matter most, at the cost of a non-commercial license. The Q8 GGUF files are 74.6 MB, the safetensors weights are 121.7 MB and there is a 60.4 MB Q4 build. The language-specific models have 12 to 37 built-in voices each and the multilingual model has 102. Voice cloning and voice design are not in the open weights and go through Gradium's studio.

Gradium's model card compares Phonon with Kokoro, Magpie and NeuTTS where those support the language, and says a Phonon model comes first on all three of its measures in every language. In English the 40M model scores an Elo of 2117 in blind listening tests with 468 raters and 8,290 votes, against 2098 for Phonon 31M, 2043 for Kokoro, 2034 for NeuTTS and 1939 for Magpie. In French, German, Spanish and Portuguese the language-specific models score 2072, 2097, 2103 and 2144, and the multilingual model scores 2016, 2046, 2076 and 2048, so a dedicated model is somewhat better than the shared one. On the card's hard set of sentences full of numbers, dates and codes, the language-specific models have word error rates of 2.1% to 2.9%, against 3.5% to 5.5% for Kokoro and 25% to 30% for Magpie where each was tested. Gradium's launch post adds a time to first audio of 8.8 ms and about 28 times real time for Phonon 40M on four threads of a MacBook Air M5 CPU, against 517 ms and about 11 times for Kokoro. All of these are the developer's own results, and we have not reproduced them.

The license is CC BY-NC, tagged as version 4.0 on Hugging Face, and Gradium points commercial users to its sales team. The repositories are gated, so you must log in and accept the terms before downloading. The runtime is a separate open-source project with backends for CPU, Metal, CUDA, Vulkan, WebGPU and Apple Core ML, and it can also run as a local server with an OpenAI-compatible API. In the 30 days to 11 November 2026, Hugging Face counted about 30,600 downloads for the German model, 27,400 for French, 21,900 for the multilingual model, 17,700 for Spanish, 12,900 for English and 11,200 for Portuguese.

Good for

  • Research and non-commercial voice projects that need small on-device speech in French, German, Spanish or Portuguese as well as English
  • One set of weights that serves five languages, with the multilingual model and its 102 built-in voices
  • Text full of phone numbers, dates and version numbers, which Gradium says is read correctly through text normalization
  • Very small files of 60 to 122 MB that run on phones, browsers and plain CPUs

Watch out for

  • CC BY-NC means no commercial use without an agreement with Gradium, and the Apache 2.0 Phonon 31M is English only
  • The repositories are gated, so you need a Hugging Face account and must accept the terms before you can download the weights
  • The model speaks only its five languages, and text normalization needs the language set to match the text
  • The Elo, word-error and speed figures are Gradium's own, measured on its own test sets and hardware

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 40M. 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-40m-multilingual (Q8 GGUF) INT8 40M 0.0746 GB 0.0858 GB Laptop or phone
phonon-40m-en (Q8 GGUF) INT8 40M 0.0746 GB 0.0858 GB Laptop or phone
phonon-40m-fr (Q8 GGUF) INT8 40M 0.0746 GB 0.0858 GB Laptop or phone
phonon-40m-de (Q8 GGUF) INT8 40M 0.0746 GB 0.0858 GB Laptop or phone
phonon-40m-es (Q8 GGUF) INT8 40M 0.0746 GB 0.0858 GB Laptop or phone
phonon-40m-pt (Q8 GGUF) INT8 40M 0.0746 GB 0.0858 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

A Creative Commons license that lets anyone copy, share and adapt the work with credit, but only for non-commercial purposes. Some model cards name it without a version number, and the current version is 4.0.

  • Attribution or model name must be shown

More about CC BY-NC

See also

Frequently asked questions

What is Phonon 40M?
Phonon 40M turns text into natural speech in five languages and is small enough for a phone, a browser tab or a laptop processor. There is one model for each language and one multilingual model that covers all five. The weights are free for non-commercial use only, and Hugging Face asks you to accept terms before you can download them.
Can I use Phonon 40M commercially?
Not without separate permission. Phonon 40M is released under the CC BY-NC, which allows research and personal use only.
Is Phonon 40M open source?
Phonon 40M is open-weight, not open source in the strict sense. You can download the model files, but the CC BY-NC adds conditions that the Open Source Initiative definition does not allow.
Can I run Phonon 40M on my own computer?
Yes. The smallest variant, phonon-40m-multilingual (Q8 GGUF), needs about 0.0858 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Phonon 40M released?
Phonon 40M 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 31M Gradium 31M – Apache 2.0 (Permissive) 0.0747 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