Sophon-Realtime v1 Puts a Laughing, Interruptible 32B Voice Model on a Single 80 GB GPU

Adventists AI has released Sophon-Realtime v1, an open full-duplex voice model family that listens while a person is still talking, decides for itself when to answer, and can laugh, sigh or breathe in its replies. The v1.0 GitHub release went out on November 8, 2026, along with a sophon-realtime 1.0.0 package on PyPI the same day. Code and weights are published under Apache-2.0, and Adventists says they are free for commercial use.

The project’s pitch is that an expressive, interruptible voice agent no longer needs a cluster. Two of the three builds run on a single 80 GB GPU through vLLM-Omni. The quickest route is pip install sophon-realtime followed by sophon-realtime serve, which downloads about 42 GB of weights on first run.

One brain, three ways to speak

All three builds share the same 32B “thinker”. It listens to the audio, decides whether the speaker has finished, and writes the reply. The model card identifies it as DuplexJev-32B-Para v1.5, stored in FP8. The builds differ only in how that reply becomes sound.

• Tag (recommended) has the thinker write inline sound tags such as [laughter], [breath], [sigh] and [cough], plus marks for stressing a word. A streaming speech model called tag-tts, based on Fun-CosyVoice3-0.5B, performs those tags while the text is still being written. The tags are never read aloud and are stripped from the text stream. Adventists reports about 0.6 seconds to first audio and up to 16 concurrent users per GPU.
• Flash speaks the same tagged reply with a lighter speech model that does one forward pass per 80 ms frame. Adventists reports 0.31 seconds to first audio and up to 40 users per GPU.
• E2E is a research preview with a talker inside the 32B model and no separate speech model. It needs two 80 GB GPUs.

The README’s advice: Tag if the voice should act, Flash if latency and users per GPU matter most.

How turn-taking works

Sophon listens continuously. At each pause, the model decides in a single token whether the user has finished, is still talking, is back-channeling (“mm-hm”) or is thinking. A reply is drafted from the first pause and dropped unheard if the user keeps talking. When a user interrupts, the model stops and keeps only the part of its answer the user actually heard in the conversation history.

Both production builds expose a WebSocket endpoint at /v1/realtime, which the project describes as OpenAI-Realtime-style, and a POST /v1/chat/completions route that takes audio in and streams text and audio out. A separate “Live mode” gateway speaks the OpenAI GPT-Live protocol, so the project says existing GPT-Live clients can switch by changing only the URL and key. In that mode Sophon keeps the voice and hands lookups and actions to a backend.

The numbers, as Adventists reports them

Sophon’s scores come from Adventists’ own runs, while some comparison scores come from outside sources. The company says the shared brain scores 82.8 on VoiceBench and 81.8% on Big Bench Audio, which it calls first among open full-duplex models, and level with GPT-4o-mini-Audio on VoiceBench. On paralinguistic understanding it reports 48.6, ahead of Gemini 3.8 Flash (46.9) and Qwen3-Omni (44.6) on its own harness. For tool use, it cites a Full-Duplex-Bench v3 Pass@1 of 48%, compared with 43% for Grok taken from the benchmark paper, using the best of several runs.

On expressiveness, the Tag build scores 0.80 in English and 0.86 in Chinese on NVV-SuperBench, a non-verbal-sound benchmark judged by Gemini 3 Flash. Flash scores 0.63 and 0.77. On one H100, median first audio for the Tag build was 0.75 seconds with 16 users and 1.36 seconds with 24, where queueing began.

Sigh and laugh tags still miss

The model card is candid about the gaps. In a 96-sentence test, a sound was detected for only 25% of written sigh tags and 42% of laugh tags, below the reference tag-tts renderer on sighs. Mandarin is slightly less clear than English with the distilled speech model. Replies longer than about a minute of speech fall outside what the voice was trained on. Only English and Mandarin are documented.

The hardware floor is also real. A Tag instance uses about 65 GB of VRAM, and Adventists says a first start takes 15 to 20 minutes while kernels and TensorRT engines build. LoRA fine-tuning in ms-swift can change persona, tone and domain knowledge, while hearing, voice and turn-taking stay as released.

The release is new: the GitHub repository had a dozen stars on November 9, and none of the benchmark claims has been independently reproduced yet.