Sophon-Realtime v1: Apache 2.0 full-duplex voice models built on a 32B brain
Adventists AI open-sourced three full-duplex voice models (Tag, Flash, E2E) sharing one 32B adapted Qwen3-VL brain that listens continuously, detects turn ends, and drafts replies from the first pause; self-reported VoiceBench 82.8.
Why it matters
A fully open, commercially usable full-duplex voice stack — the 32B brain holds a draft answer from the first pause instead of waiting for silence, and the Tag variant even emits laughter and sigh tags.
What you could build with it
A startup building a phone-based voice agent could deploy Sophon-Realtime Flash as its conversational layer, getting natural turn-taking and sub-second responses on its own GPU instead of paying per-minute for a closed voice API.
Does it hold up?
Too early to judge: all latency and benchmark figures are self-run with no independent verification, E2E is a research preview needing two GPUs, and it only understands English and Chinese.
Built with Sophon-Realtime v1
- Adventists AI releases Sophon-Realtime v1 voice family under Apache 2.0mlllm.io · Launch news: Tag, Flash and E2E variants, the shared 32B brain, and code/weights on GitHub, Hugging Face, ModelScope and PyPI.
- Adventists AI open-sources full-duplex voice models Sophon-Realtime v1mlllm.io · Deeper analysis of the three builds, the 82.8 VoiceBench claim, 0.31–0.6s latencies, and the English/Chinese-only limitation.
- adventists-ai/sophon-realtimegithub · 269 ★ · Official code and weights for the Tag, Flash and E2E full-duplex voice models, Apache 2.0.
Learn more
First spotted on hf: source.
More AI models
VoiceStudio: open-source, fully-local ElevenLabs alternative for voice cloning, dubbing, and audiobooksdebpalash/VoiceStudio is a desktop app that runs voice cloning, voice design, video dubbing, dictation, transcription…models · JEV 0.75Edge0-145B: open 145B MoE that runs on a single consumer GPUEdge0 open-sourced 145B and 120B latent-MoE models derived from Qwen3.8-Flash-Next, plus an inference framework that…models · JEV 0.74Strata: run a 125B model on a gaming PCFree, MIT-licensed local inference engine that runs Qwen3.8-Flash-Next (125B MoE, about 6B active per token) on…models · JEV 0.73Google AI Edge Gallery: offline speech recognition and synthesis on-deviceGoogle added fully offline speech recognition and text-to-speech to AI Edge Gallery, with ML Kit APIs on Android and…models · JEV 0.73Soup: fine-tune an 8B LLM on a 4 GB laptop GPU from one YAML fileAn open-source CLI that fine-tunes LLMs from a single YAML file, with layer streaming that trains an 8B model on a 4…models · JEV 0.73Tencent Hunyuan Hy3: open 295B MoE with 21B active under Apache 2.0Tencent open-sourced Hy3, a 295B-parameter mixture-of-experts model with only 21B parameters active per token, under…models · JEV 0.73
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.