Adventists AI open-sources full-duplex speech model Sophon-Realtime-v1

On November 10, 2026, Adventists AI released Sophon-Realtime-v1, a full-duplex voice model family, on Hugging Face. The release is distributed under the Apache-2.0 license and ships with an online demo, so anyone can start talking to the model in a browser without any setup.

According to the model card, talking to Sophon-Realtime-v1 is meant to feel like a phone call with a person. The model listens the whole time you talk, picks up its reply as soon as you pause, and stops immediately the moment you cut in. Rather than tracking only the words spoken, it also tracks how they are said: the model card lists hesitation, emphasis, sarcasm, laughter and sighs as cues it picks up on, and describes it answering with a tone of its own, laughing, sighing or clearing its throat when the conversation calls for it. A recommended usage tag, Flash, is listed alongside the release, and a GitHub link is provided as well.

Beyond the launch itself, the available material is thin. No parameter count, benchmark scores, training data or follow-up roadmap have been published so far, so it is not yet clear how the model stacks up against other open voice models or where it might fit in a production stack.