According to CoinMeta, the open-source ultra-small voice model Phonon has been officially released. It weighs only 64MB and supports use even when the phone is offline. The model began private testing in April of this year with approximately 100 million parameters. This time, versions with 31 million and 40 million parameters have been launched, and the model weights have been made public for the first time. The 31-million-parameter English version of q8 weighs only 64MB, while the standard CPU can run the 40-million-parameter version, which supports five languages: English, French, German, Spanish, and Portuguese, offering both single-language and five-language combined models. Gradium reveals that generating the first segment of audio on a Snapdragon 8Elite Gen 5 takes only 7.5 milliseconds, with a synthesis speed that is 24 times real-time. In terms of sound quality, the 31-million-parameter English version received a score of 2098elo, which is 3.5% lower than that of another open-source voice model with 82 million parameters, Kokoro. This model has already been integrated into tools such as Ollama and MLX-Audio. Partners Xybrid have achieved real-time dialogue for 3D game characters on a 16GBM1 Pro MacBook.