Sophon-Realtime
Adventists AI's open full-duplex voice model: a 32B thinker, three voice builds and 0.25 to 0.6 s to first audio on one 80 GB GPU.
In short. Sophon-Realtime is a voice assistant model you run yourself. It listens while you talk, works out when to jump in, stops when you interrupt, and can laugh or sigh in its replies. It needs one large 80 GB graphics card, and the weights are free under Apache 2.0.
Summary
Sophon-Realtime v1 is a family of open full-duplex voice models from Adventists AI. A 32B thinker listens to the caller's audio the whole time and, at each pause, decides in a single token whether the speaker has finished, is still talking, is just back-channeling or is thinking. It drafts a reply from the first pause and writes it with inline sound tags such as a laugh or a sigh. Three builds share this thinker and differ only in how the reply becomes sound: Tag uses an expressive streaming voice model, Flash a lighter and faster one, and E2E a compact talker of about 350M parameters that generates continuous voice vectors with no speech tokens. The thinker builds on Adventists AI's own DuplexJev model, whose card lists Qwen3-VL-32B-Instruct and Qwen3-ASR-0.6B as its bases.
The developer reports a VoiceBench score of 82.8, level with GPT-4o-mini-Audio and ahead of MiniCPM-o 4.5 at 76.1, and 81.8% on Big Bench Audio against 90% for GPT-Live-1. On Full-Duplex-Bench v3, a tool-calling test, it reports a Pass@1 of 48%, the best of several runs, below GPT-Realtime at 60% and Gemini Live 3.1 at 54% and above Grok at 43%. On NVV-SuperBench, which checks whether the laughs and sighs a reply calls for really appear in the audio, the Tag build scores 0.86 in Chinese and 0.80 in English. On one H100 80 GB, it reports first audio after about 0.6 s for Tag, 0.31 s for Flash and 0.25 s for E2E, and recommends up to 40 concurrent users per GPU for Flash and 16 for Tag and E2E. Adventists AI says its benchmarks are self-run unless noted, and we have not reproduced any of these numbers.
Code and weights use Apache 2.0, and each build's NOTICE file lists third-party components that keep their own licenses. Tag and Flash are served with vLLM-Omni through a WebSocket realtime API and a chat-completions endpoint, and a gateway speaks the OpenAI GPT-Live protocol. Version 1.1, released on 11 November 2026, adds zero-shot voice cloning from a 3 to 10 second clip in Tag and Flash. Chinese and English are fully supported, the project's own tests call Italian, German, Spanish, French and Portuguese ready for conversation, and Flash adds Japanese and Russian, while E2E speaks Chinese and English with one built-in voice. The first public release came on 8 November 2026, and in the 30 days to 11 November 2026 Hugging Face counted about 440 downloads for the Tag weights and 310 for Flash.
Good for
- Voice agents that need real full-duplex behavior, with barge-in and back-channels, on one 80 GB GPU you control
- Companion, character and demo voices that laugh and sigh, with the Tag build
- Customer-service or in-car deployments with many callers per GPU, with the Flash build
- Teams that want an Apache 2.0 voice stack and will fine-tune persona, tone and tool use with LoRA in ms-swift
Watch out for
- It needs a data-center class GPU, one 80 GB card per instance on Linux with CUDA, and the first start of Tag takes 15 to 20 minutes
- Most benchmark figures are self-run by the developer, the tool-calling score is the best of several runs, and GPT-Realtime and Gemini Live score higher there
- The project is days old, with versions changing fast, and has had few downloads so far
- Voice cloning is in Tag and Flash only, and the developer asks users to clone only their own voice or one they have permission to use
- The Sophon builds are Apache 2.0 while the separately published DuplexJev models are CC BY-NC, so read each build's NOTICE file for third-party terms
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Sophon-Realtime. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Sophon-Realtime-32B-Tag | 32B | 42 GB | 80 GB | Workstation |
| Sophon-Realtime-32B-Flash | 32B | – | 80 GB | Workstation |
| Sophon-Realtime-32B-E2E | 32B | – | 80 GB | Workstation |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 80 GB of GPU memory.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is Sophon-Realtime?
- Sophon-Realtime is a voice assistant model you run yourself. It listens while you talk, works out when to jump in, stops when you interrupt, and can laugh or sigh in its replies. It needs one large 80 GB graphics card, and the weights are free under Apache 2.0.
- Can I use Sophon-Realtime commercially?
- Yes. Sophon-Realtime is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Sophon-Realtime open source?
- Sophon-Realtime is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Sophon-Realtime on my own computer?
- Only on a workstation. The smallest variant, Sophon-Realtime-32B-Tag, needs about 80 GB of memory at 4-bit precision, so plan for a 48 to 96 GB GPU, several graphics cards, or a Mac with plenty of memory.
- When was Sophon-Realtime released?
- Sophon-Realtime was released by Adventists AI on November 8, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Sophon-Realtime v1 family page with benchmark tables (Hugging Face) huggingface.co
- Sophon-Realtime-32B-Tag model card (Hugging Face) huggingface.co
- Sophon-Realtime-32B-Flash model card (Hugging Face) huggingface.co
- Sophon-Realtime-32B-E2E model card (Hugging Face) huggingface.co
- Sophon-Realtime README, languages and changelog (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| LFM3-Audio-2B Liquid AI | 2B | 32K | LFM Open 1.0 (Conditional) | 4.6 GB Laptop or phone | |
| Voxtral Realtime Mistral AI | 4.43B | 128K | Apache 2.0 (Permissive) | 10 GB Laptop or phone | |
| Qwen3-ASR Alibaba Qwen | 918M – 2.35B | – | Apache 2.0 (Permissive) | 2.12 GB Laptop or phone | |
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone |