Sophon-Realtime-v1: 32B open-source model with GPT-4o-mini-level voice conversation

A full-duplex voice model that responds instantly the moment a user stops speaking and stops immediately if interrupted. A single 32B parameter model handles listening, turn-taking, and response generation simultaneously. Released under the Apache-2.0 license, it permits commercial use.

Going beyond simple text conversion, it understands the counterpart's tone and emotions. It recognizes paralinguistic elements such as onomatopoeia, emphasis, sarcasm, laughter, and sighs, and naturally includes non-verbal sounds like laughter or coughing in its responses. It achieved top-tier performance among open-source full-duplex models with a VoiceBench score of 82.8 and 81.8% on Big Bench Audio.

The Tag, Flash, and E2E builds share the same brain, differing only in output method. Tag offers expressiveness, Flash provides a speed of 40 concurrent users per 80GB GPU, and E2E delivers first audio output within 0.25 seconds. It is compatible with the OpenAI GPT-Live protocol, allowing existing clients to be used as-is.

It is suitable for environments requiring real-time voice interfaces, such as customer service, in-car systems, and character-based chatbots. Deployment is possible with a single command using vLLM-Omni, and domain knowledge or tone can be fine-tuned via LoRA. Every response includes a text transcript for gratitude and tool integration.