Gradium released Phonon speech synthesis models that can run on phones, laptops, and in browsers. The English 31M model has approximately 64MB of quantized weights and is available under Apache 2.0. The 40M individual-language and multilingual models support five languages, including English and French, but their CC BY-NC license has different conditions for commercial use. The company measured roughly 7ms to first audio on a MacBook Air M5 CPU. That figure comes from specific hardware and settings. Beyond the benefits of local processing, product teams should separately check language support, audio quality, and licensing.