Skip to content
Neural Daily

ModelsNews

Sophon-Realtime v1: open full-duplex voice on one GPU

🇭🇺 Olvasd el magyarul →

Adventists AI released the Sophon-Realtime v1 voice model family under Apache 2.0. The 32B model listens continuously, decides when to speak, picks up tone of voice and can add a laugh or a sigh to its reply. The Tag and Flash builds run on a single 80 GB GPU.

· 3 min read

A Sophon-Realtime-32B-Tag modell Hugging Face-oldalának előnézeti képe
Image: Hugging Face · Screenshot · Source

What was released

Adventists AI published Sophon-Realtime v1 on November 8, a day ahead of its announced date. It is a full-duplex speech-to-speech family: the model keeps listening while it talks, starts drafting while the user is still speaking, stops when interrupted and keeps going on a simple "mm-hm". Code and weights are Apache 2.0 (Hugging Face).

There are three builds on the same 32B brain, differing only in how the reply becomes sound:

  • Tag (recommended): expressive voice, about 0.6 s to first audio, 16 users per GPU.
  • Flash: lighter, faster voice, 0.31 s to first audio, 40 users per GPU.
  • E2E (preview): built-in talker with no separate speech model, about 1 s, needs two 80 GB GPUs.

It hears tone and answers with tone

The model picks up hesitation, emphasis, sarcasm, laughter and sighs from speech. In its own runs it scores 48.6 on paralinguistic understanding, against 46.9 for Gemini 3.8 Flash and 44.6 for Qwen3-Omni. It writes inline sound tags such as a laugh or a sigh into its reply, and the voice performs them instead of reading them aloud.

Other figures: VoiceBench 82.8, Big Bench Audio 81.8% and Pass@1 48% on Full-Duplex-Bench v3 tool calling. All are self-reported; there is no independent evaluation yet.

How to run it

You need Linux, CUDA, Python 3.12 and one 80 GB GPU (H100, H800 or A100 80G), about 65 GB of VRAM per instance. Serving is based on vLLM-Omni. There are two APIs: a WebSocket for full duplex and an OpenAI-style chat completions endpoint. An optional Live mode gateway speaks OpenAI's GPT-Live protocol, so existing clients can switch by changing only the URL and key (GitHub).

Persona, tone and domain knowledge can be fine-tuned with LoRA in ms-swift on one GPU. Hearing, voice and turn-taking stay as released.

Caveats

The model card says sighs and laughing while speaking are still weak: the model writes those tags more often than the voice delivers them. Mandarin is slightly less clear than English, and replies longer than about a minute go beyond what the voice was trained on.

If you build voice agents and want to self-host, this is one of the most serious open options right now. On the closed side, Eleven Realtime and GPT-Live are the reference points (Eleven Realtime: speech-to-speech models for voice agents).

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles