✦
Summary
Sophon-Realtime is an open full-duplex voice model family that hears tone, not just words: it picks up emotion, emphasis, sarcasm, laughter and sighs, and answers with an expressive voice that laughs, sighs and breathes. Three builds share one 32B brain: Tag for expressive speech, Flash for speed and scale at 40 users per GPU, and E2E for the fastest first audio at about 0.25 seconds. New 4B builds run on a single consumer GPU with 24 to 32 GB of memory. The code and weights are Apache 2.0 and free for commercial use.
Key Features
Full-duplex conversation with barge-in: listens continuously and stops when interrupted
Paralinguistic understanding: hears emotion, emphasis, sarcasm, laughter and sighs
Expressive speech with non-verbal sounds such as laughter, sighs and breathing via sound tags
Zero-shot voice cloning from a 3 to 10 second audio clip
Three builds: Tag for expressive speech, Flash for speed and scale, E2E for fastest first audio around 0.25 seconds
4B builds for a single consumer GPU with 24 to 32 GB of memory
Offline Mac and Apple-silicon run via MLX, or self-hosting on one 80 GB GPU via vLLM-Omni
Tool calling and an OpenAI Realtime and GPT-Live-compatible Live mode
About Sophon-Realtime
The model clones a voice zero-shot from a 3 to 10 second clip, and supports tool calling with a Full-Duplex-Bench v3 Pass@1 score of 48%. It runs self-hosted on one 80 GB GPU via vLLM-Omni with a single command, fully offline on a Mac with MLX, or through a hosted OpenAI-Realtime-compatible API that offers 3 hours of free voice chat as a trial. Multilingual support covers Chinese and English plus Italian, German, Spanish, French and Portuguese, with Japanese and Russian on Flash.
Sophon-Realtime is developed by adventists-ai and had 413 stars on GitHub shortly after release, with version 1.4 released in November 2026. It is aimed at developers building voice agents that need to sound human, hear how a caller feels, and act on what is said, without paying for a closed voice API.