🤖 Adventists AI releases Sophon-Realtime v1 voice models
On November 8, 2026, code and weights were released on GitHub, Hugging Face, ModelScope, and PyPI. The Tag, Flash, and E2E builds share a single 32B 'brain' (adapted Qwen3-VL-32B-Instruct): it listens, decides when the interlocutor has finished, and holds a draft response from the first pause; only the voice output differs.
🌍 According to the authors' measurements, this is the first open full-duplex model at the level of closed ones: VoiceBench 82.8, similar to GPT-4o-mini-Audio, and Pass@1 48% compared to Grok's 43% (Full-Duplex-Bench v3). Apache 2.0 permits commercial use; the figures are self-run, with no independent verification.
👤 The model can be installed on a single 80-GB GPU via the command pip install sophon-realtime (weights ~42 GB): Tag has ~0.6 s to first audio and supports 16 users per GPU, while Flash has 0.31 s and supports up to 40 users.
Source 1: https://huggingface.co/adventists-ai/Sophon-Realtime-v1 Source 2: https://github.com/adventists-ai/sophon-realtime