Sophon-Realtime v1: Open Full-Duplex Voice AI on One GPU

Sophon-Realtime v1 is an open full-duplex voice AI model released on 8 November 2026 by Adventists AI. It listens and speaks at the same time, like a real phone call: you can interrupt it, and it stops the moment you cut in. The code and weights are Apache 2.0, so a business may use them commercially and self-host the model on a single 80 GB GPU instead of paying a voice provider per minute. All three builds share one 32-billion-parameter brain adapted from Qwen3-VL-32B-Instruct, which listens continuously, decides when a turn ends and writes the reply.

What Sophon-Realtime v1 actually is

Sophon-Realtime v1 is a family of three voice models, released together on GitHub, Hugging Face, ModelScope and PyPI under Apache 2.0 — so a business can run it for its own customers without per-minute API fees.

One 32B brain handles listening, turn-taking decisions and reply writing; the three builds differ only in how the reply becomes sound. The hard part of a voice agent is knowing when to speak, wait and stop — and this release puts that judgement in the model, not hand-written rules. Our guide to how to build an AI voice agent covers the same pipeline: listen, think, speak.

Full duplex, in plain English

Most voice agents are half-duplex: you speak, the agent waits for silence, then it answers — like a walkie-talkie. A real phone call is full duplex: both sides can speak at once, you can interrupt, and the other person can say "mm-hmm" while you talk. That is what Sophon-Realtime v1 aims for — see our coverage of Block's Dessa and Meta's Muse Voice.

How the brain decides when to speak

At each pause the brain decides in one token whether you have finished, are still talking or are thinking, and drafts a reply from the first pause, dropping it if you keep talking. If you interrupt, it stops at once and keeps only the part you actually heard. It also marks laughter, sighs and emphasis for the voice to perform.

Why callers will notice the difference

Interruptibility is the test callers feel: a half-duplex agent that talks over a correction forces the caller to wait and repeat themselves, while one that stops and adjusts feels like a person. Reception and support are the natural first uses, as our guide to AI receptionists for small business explains.

The three builds: Tag, Flash and E2E

The three builds share one brain but serve different jobs:

• Tag (recommended, most expressive): about 0.6 seconds to first audio, around 16 concurrent users per GPU. The voice, derived from Fun-CosyVoice3-0.5B, performs the expressive marks the brain writes in. Best for companion-style calls.
• Flash (fastest, highest concurrency): first audio in 0.31 seconds, up to 40 concurrent users per GPU, using a lighter voice with one forward pass per 80 ms frame. Best for customer service, bookings and in-car use.
• E2E (research preview): puts the talker inside the 32B model, at about 1 second to first audio — but needs two 80 GB GPUs. Treat it as a direction of travel, not a deployment target.

All latency and concurrency figures are vendor-reported, measured on one H100. Test Flash first for support lines, Tag when the voice needs warmth.

The benchmark scores — read them as vendor claims

The headlines are VoiceBench 82.8 and Big Bench Audio 81.8%, with the team claiming first place among open full-duplex models and parity with GPT-4o-mini-Audio on VoiceBench. Those are strong numbers, but they are self-run by the team that built the model, with no independent evaluation yet — so treat them as vendor-reported, as we do for Phonic-1 and Nemotron Transcribe.

What the supporting scores add

Tone understanding is 48.6, ahead of Gemini 3.8 Flash (46.9) and Qwen3-Omni (44.6); tool calling is 48% Pass@1 versus 43% for Grok. Tone matters because an agent that cannot hear frustration keeps repeating the wrong answer; tool calling matters because one that cannot check a booking is just a talking FAQ. The team also lists its weak spots: laugh-while-speaking tags are still weak, Mandarin is slightly less clear than English, and replies over about a minute exceed the training range.

What it takes to run it

You need Linux, CUDA, Python 3.12 and one 80 GB GPU (H100, H800 or A100 80G) for Tag and Flash. The fastest path is pip install sophon-realtime then sophon-realtime serve, which downloads the roughly 42 GB Tag weights on first run and serves an OpenAI Realtime-style WebSocket on port 8000. Its Live mode gateway speaks the OpenAI GPT-Live protocol, so existing GPT-Live clients only change the URL and key.

That is a real server, not a laptop job: a rented 80 GB GPU is a meaningful monthly cost, and someone must keep it patched. Price GPU rental, engineering time and monitoring against a per-minute hosted API. Our guide to AI voice calling agents sets out the same comparison, and our custom software cost guide for the UK shows how to budget the build. Check the Sophon-Realtime repository on GitHub for the current requirements before you start.

Where it fits in a real business

It fits the calls businesses already pay people to handle, where the caller expects a conversation, not a menu:

• Reception and first contact: answering, routing and taking messages without a phone tree.
• Support triage: gathering the facts and booking the human follow-up, with tool calling to pull real account data.
• Bookings and reservations: a voice agent that can be interrupted mid-sentence to correct a date is useful — our work on restaurant systems shows how much of that trade still runs on calls.
• Outbound reminders: appointment reminders and delivery windows, where the call is short and the script is narrow.

It does not fit everywhere. Long advisory calls exceed the roughly one-minute reply range, and high-stakes calls need a human in the loop. Calls handling health or financial detail need the care our guide to AI chatbots for clinics describes: consent, data handling and a clear escalation path.

Self-host versus rent per minute

The comparison is fixed cost versus variable cost. A hosted API costs nothing when the phone is quiet and scales with every minute of talk. Self-hosting costs a GPU whether anyone calls or not — but the thousandth call costs close to nothing, and the model, data and recordings stay on your infrastructure: no audio leaves your servers.

Volume decides: a clinic taking a few dozen calls a day will find a hosted API cheaper and simpler, while a support line taking thousands of minutes a day, or a business with strict data rules, is where self-hosting pays. The model is days old, so expect updates, not stability.

What your business should do now

Do not rebuild your phone system around a model released yesterday. Instead, run a contained test: spin up Flash on a rented GPU, point it at one narrow job — appointment confirmations — and run it beside your existing line for two weeks. Score it like a new hire: did it understand the caller, interrupt correctly and hand over cleanly? Test it on your own recordings, not the vendor's demos.

Then compare it against the hosted options, including the chatbot-versus-agent choice and our AI calling agent product work — our GDPx and LongCat-2.5 coverage shows how fast this field is moving. If the test wins, you have an open, self-hosted voice layer with no per-minute bill. If you want help scoping that pilot, talk to Coheser about an AI agent pilot or contact us.

Frequently asked questions

What is Sophon-Realtime v1?

Three open full-duplex voice AI models released on 8 November 2026 by Adventists AI, sharing one 32B brain adapted from Qwen3-VL-32B-Instruct, under Apache 2.0 for commercial use.

Can a business use Sophon-Realtime v1 commercially?

Yes — code and weights are Apache 2.0, which allows commercial use including self-hosting. You still need your own GPU and remain responsible for its behaviour on live calls.

Which build should I try first?

Flash: fastest at 0.31 seconds to first audio, up to 40 users per GPU, suiting support and booking lines. Choose Tag for warmth and character; E2E is a research preview needing two 80 GB GPUs.

Are the VoiceBench 82.8 and Big Bench Audio 81.8% scores independently verified?

No — all figures are self-run by the team, so they are vendor-reported. Test the model on your own calls first.