Gradium is open-sourcing Phonon, on-device text-to-speech for apps, browsers, and small boards that need speech without a network call or a per-request bill. The models follow a private beta, started in April 2026, of a larger Phonon checkpoint. Feedback from that beta shaped this release: smaller weights, five languages, and voices that can be cloned or designed in Gradium Studio and then run locally. Phonon 31m is English only, 31 million parameters, 64 MB at q8, under Apache 2.0, with 37 built-in voices. Phonon 40m is 40 million parameters and 74.6 MB, either one model each for English, French, German, Spanish, and Portuguese, or one multilingual model with 102 voices. The 40m weights are CC BY-NC. On a public hard-sentence set of 500 cases per language (phone numbers, emails, codes), Gradium reports a 1.70% word error rate for Phonon 31m, against 3.50% for Kokoro and 8.71% for NeuTTS Nano on the same test. A Rust runtime with no PyTorch dependency covers CPU, CUDA, Vulkan, Metal, WebGPU, Apple Core ML, Snapdragon NPUs, and the browser. On a MacBook Air M5 with Apple Accelerate and KleidiAI, Gradium measured 5 ms to first audio for Phonon 31m on one core, at 26 times real time. Arm measured 28 ms to first audio and 15 times real time with KleidiAI. Weights are on Hugging Face. The xn-ptts repo ships Python, Swift, browser, Android, and an OpenAI-compatible server. Gradium, the Paris voice company founded by Neil Zeghidour with Olivier Teboul, Laurent Mazaré, and Alexandre Défossez after their work at Kyutai, already sells cloud TTS, speech-to-text, and voice design. Phonon is the open local counterpart.