BK Jin · Adventists AI

Founder @ Adventists AI. Building open voice models that hear how you speak — DuplexJev · Sophon-Realtime (open weights soon). Melbourne.

[BK Jin · Adventists AI @BKJinAI]

No GPU? Try Sophon-Realtime in the cloud.

Sign up with an email and get 3 hours of free full-duplex voice chat.
• Tag, Flash and E2E models
• OpenAI Realtime-compatible: point your SDK at it
• Barge-in, back-channels, English & Chinese
https://t.co/0hEZIIWWiL
@ModelScope2022

[BK Jin · Adventists AI @BKJinAI]

@heybackchannel @_MarcosValera @RealRichomie Thanks for including us! Two things from this week:
• Sophon-Realtime now runs fully offline on a Mac (MLX)
• No GPU? Our hosted API is OpenAI Realtime-compatible, and sign-up comes with 3 hours of free voice chat

https://t.co/0hEZIIWWiL

[BK Jin · Adventists AI @BKJinAI]

Full-duplex voice AI, fully offline on one Mac.

Sophon-Realtime now runs on Apple silicon (MLX). No cloud, no internet.
• Hears your voice directly, no speech-to-text
• Cut in anytime: stops in ~0.1 s
• ~1 s to reply on an M4 Max

https://t.co/a9YmQPMmG5
@ModelScope2022

[BK Jin · Adventists AI @BKJinAI]

@ModelScope2022 Thanks @ModelScope2022! The studio now runs v1.2: long replies are spoken to the end.

[BK Jin · Adventists AI @BKJinAI]

Sophon-Realtime v1.2 is out.
• Long replies are now spoken to the end, no cap on speech length.
• New: Flash-4bit — full-duplex voice chat on one 32 GB GPU, ~0.2 s first audio, barge-in + tool calls.
https://t.co/oUqnkMsEmI
https://t.co/nIJ6njampP

[BK Jin · Adventists AI @BKJinAI]

@ModelScope2022 Next up: the E2E build with 0.21 s to first audio, later this week. Single consumer GPU and Mac deployment around Nov 20.

Please only clone your own voice or voices you have permission to use.

[BK Jin · Adventists AI @BKJinAI]

Sophon-Realtime v1.1: zero-shot voice cloning in our open full-duplex model.

A few seconds of audio and it talks in that voice, live. No fine-tuning, no restart, one GPU.

Video: my voice from 11 s, then a colleague's from 9 s. Thanks @ModelScope2022!
https://t.co/Yq1MzsPB36

[ModelScope @ModelScope2022]

Sophon-Realtime v1.1 from Adventists AI @BKJinAI adds zero-shot voice cloning to an open-weights, full-duplex voice model.

Drop a few seconds of someone's voice into model/voices/, set "voice" to its name, and talk to it in that voice in real time. No fine-tuning, no restart, one GPU. Barge-in and laughs still work, and a voice cloned from a clip in one language can speak another.

Available in the Tag and Flash builds on ModelScope. Live mode is a drop-in replacement for the OpenAI GPT-Live API, and an E2E build with 0.21 s to first audio is coming soon.

Please only clone voices you have permission to use. 🤖https://t.co/jkFd0JRhhW

[BK Jin · Adventists AI @BKJinAI]

@ModelScope2022 Also in the video: Live mode. Sophon hands lookups like weather or news to your own backend, keeps talking, then says the answer. It's a drop-in for the OpenAI GPT-Live API, and we'll open-source the backend framework too.

[BK Jin · Adventists AI @BKJinAI]

@TrelisResearch Update: everything in your report plus long-reply edge cases is fixed in 1.1.2 (pip install -U "sophon-realtime[trt]"). Clean re-run of your protocol: 32 long sessions, 0 errors. 1.1 also adds zero-shot voice cloning + Live mode. Would love a re-benchmark! https://t.co/e1K4Ayembb

[BK Jin · Adventists AI @BKJinAI]

Sophon-Realtime v1 is out: an open full-duplex voice model that hears your tone and speaks with one.

It listens while you talk, catches sarcasm and stress, and laughs or sighs when the moment calls for it.

Single live takes, no edits, no speed-up. Open weights, Apache-2.0.

[BK Jin · Adventists AI @BKJinAI]

@TrelisResearch Thanks for the very thorough report! Fixed in sophon-realtime 1.0.2 (pip install -U sophon-realtime): the first realtime session after a cold start could lose its reply. Details and verification in the HF thread. Would love to see your benchmark!

[BK Jin · Adventists AI @BKJinAI]

@nikhilgoyalai Fair take, thanks for the write-up! Today Tag and Flash need one 80 GB GPU. Next up: a quantized build aimed at a single 24 GB consumer GPU, and an Apple MLX build for Macs later this month. Meanwhile you can try it in the browser: https://t.co/x3H1LrXnXR

[BK Jin · Adventists AI @BKJinAI]

@Forhanvv Thanks for sharing! Author here. You can talk to it in the browser: https://t.co/r8SSgeduS1 · code: https://t.co/nIJ6njampP

[BK Jin · Adventists AI @BKJinAI]

Thanks @ModelScope2022 for the feature! Sophon-Realtime v1 is open on ModelScope: it hears your tone and speaks with one.

Talk to it in your browser: https://t.co/r8SSgeduS1

Or self-host on one GPU with vLLM-Omni:
pip install sophon-realtime && sophon-realtime serve

[ModelScope @ModelScope2022]

Adventists AI just open-sourced Sophon-Realtime v1 on ModelScope: a full-duplex voice model that hears how you speak, and answers with tone of its own.

• It keeps listening while you talk, answers ~0.4 s after you stop, and stops the moment you cut in
• It laughs, sighs and clears its throat when the conversation calls for it
• #1 among open full-duplex models on VoiceBench (82.8) and Big Bench Audio (81.8%)
• 48% Pass@1 on Full-Duplex-Bench v3 tool calling, ahead of Grok (43%)
Apache-2.0 weights. Built by @BKJinAI and team.
https://t.co/a4vfp6R34x
https://t.co/EVJesDqKUg

[BK Jin · Adventists AI @BKJinAI]

@nahid_pro09 Author here, thanks for sharing! It isn't the Gemma model from the quoted post: Sophon-Realtime is our own 32B voice model (Qwen3-32B base).

Talk to it in the browser: https://t.co/r8SSgeduS1
Code: https://t.co/nIJ6njampP

[BK Jin · Adventists AI @BKJinAI]

Try it in your browser, nothing to install. Thanks @ModelScope2022 for featuring it on the ModelScope homepage.

https://t.co/r8SSgeduS1

[BK Jin · Adventists AI @BKJinAI]

Two builds, each on one GPU with vLLM-Omni:
· Tag: the most expressive voice, ~0.6 s to first audio
· Flash: 0.31 s to first audio, 40 users per GPU

pip install sophon-realtime

https://t.co/nIJ6njampP

[BK Jin · Adventists AI @BKJinAI]

A transcript keeps the words and drops the rest: the anger in "wait, wait, stop", a laugh with no words at all, which name got the stress.

That's what a voice agent needs before it decides to talk. 30 s of what DuplexJev hears:
https://t.co/t3ivelChPf

[BK Jin · Adventists AI @BKJinAI]

On 1,035 paralinguistic test items (LISTEN, StressTest, MSPB) it scores 50.2, ahead of Gemini 3.8 Flash (47.8), Qwen3-Omni (45.0) and GPT-audio (40.9).

Open weights: https://t.co/XmjXJJtAty

[BK Jin · Adventists AI @BKJinAI]

Most voice agents decide when to talk with a silence timer. We think they should listen.

DuplexJev hears turn-taking, interruptions, emotion, stress and laughter straight from the audio: 10 decisions in 0.09 s.