Adventists AI released the Sophon-Realtime v1 family of full-duplex voice models on 8 November, publishing code and weights under the Apache 2.0 licence on GitHub, Hugging Face, ModelScope and PyPI. All three builds share one 32B "brain" adapted from Qwen3-VL-32B-Instruct that listens continuously, decides when to speak and writes the reply; they differ only in how that reply becomes sound.
The fastest build, Flash, reaches first audio in 0.31 seconds and serves up to 40 concurrent users on a single 80 GB GPU. The end-to-end build is a research preview that needs two 80 GB GPUs. The voice itself derives from Fun-CosyVoice3-0.5B, and the Apache 2.0 licence covers both code and weights, so the models can be used commercially.
On the team's own benchmarks, Sophon-Realtime v1 scores 82.8 on VoiceBench and 81.8% on Big Bench Audio โ first among open full-duplex models, and level with GPT-4o-mini-Audio on VoiceBench. Those figures are vendor-reported and self-run by the team, not independently verified, so treat them as a reason to test the models rather than as settled results.
Getting started is a one-liner: pip install sophon-realtime, then sophon-realtime serve downloads the roughly 42 GB Tag weights on first run and serves an OpenAI Realtime-style WebSocket on port 8000, so existing Realtime-style clients can point at it with minimal changes.
The release matters because full-duplex voice โ listening and speaking at the same time, with natural interruptions โ has mostly lived behind closed APIs. An open-weight family that runs on a single 80 GB GPU gives developers a self-hosted alternative for voice agents, call handling and realtime assistants, with the weights, training recipe and benchmarks all inspectable.