Introduction

Adventists AI released Sophon-Realtime v1, an open full-duplex speech-to-speech voice model family, on November 8. The v1.0 GitHub release ships two production builds, Sophon-Realtime-32B-Tag and Sophon-Realtime-32B-Flash, plus a research-preview E2E build. Code and weights are Apache-2.0. You can pull the weights from Hugging Face or ModelScope today, serve them with pip install sophon-realtime, or just talk to it in the ModelScope Studio demo.

The team had been teasing a November 9 open-source date. The release went out a day early.

Why it's interesting for voice makers

Most voice agents still take turns like walkie-talkies. Sophon listens the whole time you talk, starts drafting while you're still speaking, picks up as soon as you stop, and stops the moment you cut in. A plain "mm-hm" doesn't knock it off its line.

The part I care about most is paralinguistics in both directions:

  • It hears tone. Adventists says the shared 32B "brain" catches hesitation, emphasis, sarcasm, laughter and sighs. It scores 48.6 on their paralinguistic-understanding mix (LISTEN, StressTest, MSPB), against 46.9 for Gemini 3.8 Flash and 44.6 for Qwen3-Omni in their own runs.
  • It answers with tone. The model writes its reply with inline sound tags, and the voice performs them: a laugh after a joke, a sigh for bad news, stress on the word that matters. It doesn't read the tags aloud. On NVV-SuperBench the Tag build scores 0.86 (Chinese) / 0.80 (English).

Tag, Flash or E2E?

All three builds share the same 32B brain, which listens, decides turns and writes the reply. They differ only in how that reply becomes sound. All numbers below are vendor-reported, measured on one H100 80 GB:

  • Tag (recommended): expressive tag-tts voice (built on Fun-CosyVoice3-0.5B). About 0.6 s to first audio, up to 16 users per 80 GB GPU, one GPU, about 42 GB download and about 65 GB of VRAM per instance.
  • Flash: a lighter, faster voice for scale. 0.31 s to first audio, up to 40 users per GPU, one GPU. NVV 0.77 zh / 0.63 en.
  • E2E (research preview): a talker inside the 32B model with no separate speech model. About 1 s to first audio, needs two 80 GB GPUs. Adventists says a newer build is in testing.

Shared numbers: VoiceBench 82.8 and Big Bench Audio 81.8%, which the team calls the strongest brain among open full-duplex models. Full-Duplex-Bench v3 tool calling is reported at Pass@1 48%.

Under the hood, the Tag card lists the thinker as DuplexJev-32B-Para v1.5 in FP8. It ships six reference voices: chelsie (the default; chelsie_zh / chelsie_en picked by language), vivian_zh, ryan_en, brand_zh and brand_en.

Run it

You need Linux, CUDA, Python 3.12 and one 80 GB GPU (H100, H800 or A100 80G). The quickest route is the Tag build:

pip install sophon-realtime
sophon-realtime serve   # downloads Sophon-Realtime-32B-Tag (~42 GB) on first run, serves on port 8000

Or use the isolated one-command scripts:

git clone https://github.com/adventists-ai/sophon-realtime && cd sophon-realtime/tag
bash install.sh
GPU=0 PORT=8000 bash start.sh
python client_example.py --wav my_question.wav --lang en --voice ryan_en

Both builds expose ws://HOST:8000/v1/realtime for full duplex with turn-taking, barge-in and backchannels, plus POST /v1/chat/completions (audio in, text and audio out). There's also an optional Live mode gateway that speaks the OpenAI GPT-Live protocol. Existing GPT-Live clients can switch over by changing only the URL and key:

cd sophon-realtime/live && BUILD=tag PORT=8000 LIVE_PORT=8780 bash start_live.sh

To change persona, tone or domain knowledge, the repo includes a LoRA recipe for ms-swift on one 80 GB GPU. Hearing, voice and turn-taking stay as released.

Honest caveats

  • Every benchmark here comes from Adventists' own runs and is self-reported. Give it your own ears before you trust a leaderboard.
  • The card says [sigh] and the laugh-while-speaking tags are still weak: the brain writes them more often than the voice delivers them.
  • Mandarin is slightly less clear than English on the distilled voice, and replies longer than about a minute of speech go past what the voice was trained on.
  • This isn't laptop-sized. You need one 80 GB card per Tag or Flash instance, or two for E2E.

Conclusion

If you're building companions, characters or call-center agents and want something open that can actually laugh at your joke mid-sentence, this is the one to try this weekend.

—Anabel ♡