Gradium released Phonon as open weights on October 20. It is a family of small text-to-speech models built to run on the device rather than on a server. The model card says Phonon "builds on Pocket TTS and the Mimi codec, developed by Kyutai." The smallest version, Phonon 31m, is English only and licensed under Apache 2.0.

What was released

Gradium lists three versions:

  • Phonon 31m: English only, Apache 2.0.
  • Phonon 40m: one model per language for English, French, German, Spanish and Portuguese, with 12 to 37 built-in voices per language, under CC BY-NC.
  • Phonon 40m multilingual: the same five languages in one set of weights, under CC BY-NC.

CC BY-NC means non-commercial use only. Gradium asks companies to contact it for commercial use of the 40m models. Only Phonon 31m can be used freely in commercial products.

What 64.5 MB actually means

The headline 64.5 MB is the size of one file: model.q8.gguf, the 8-bit quantized weights of the 31m English model. It is not the total memory needed at runtime, and Gradium has not published a runtime memory figure. The 31M parameter count also excludes the audio tokenizer, the Mimi codec that turns the model's output into sound. The 40m versions are listed at 74.6 MB.

How it runs

The model card says Phonon streams audio while it is still generating, and is small enough to run in a browser tab, on a phone or on a laptop CPU. Gradium provides Python and browser paths plus a runtime on GitHub. It says the models are optimized for more than 10 hardware platforms across CPU (Intel, Arm, WebAssembly), GPU (CUDA, Vulkan, Metal, WebGPU) and NPU (Apple Core ML, Snapdragon).

Gradium's own numbers

All of the figures below come from Gradium's own testing, and we have not tested Phonon:

  • Phonon 31m starts speaking in 7 ms on a laptop CPU, Gradium says.
  • In blind pairwise listening tests against Kokoro, Magpie, NeuTTS and Pocket TTS, Gradium reports 1,189 raters and 17,270 votes.
  • On Gradium's own set of 500 hard sentences per language, word error rate was 1.70% for Phonon 31m, 2.11% for 40m multilingual and 2.22% for 40m.

Offline TTS or a cloud voice API

An on-device model like Phonon 31m makes sense when:

  • The text should not leave the device, for privacy or compliance reasons.
  • The app has to keep talking without a network, such as in cars, field tools or offline games.
  • Usage is high enough that per-character API charges add up, since local inference has no per-request fee.
  • Latency matters and you want to avoid a network round trip.

A cloud voice API is still the easier pick when:

  • You need languages beyond the five Phonon covers, or commercial use of anything other than English 31m.
  • You want large, high-end voices, voice cloning services or a managed SLA.
  • Your target devices are too constrained, or you would rather not ship and update model files yourself.

For a look at a cloud-side option, see our coverage of Grok Voice 3.0. For other small models built for local use, see MiniCPM-V 4.7 on phones and how Perplexity Computer runs locally on Mac.