Sophon-Realtime-v1: an open model of full voice dialogue with understanding of sarcasm and laughter The architecture is built on the basis of a 32 billion language model, and implements honest full-duplex dialogue (fully complex) speech-to-skeleton: the system continuously listens to the user during his speech, immediately uploads a copy when he stops and immediately falls silent (barge-in) if a person interrupts. The main technological feature of Sophon is the natural perception of non-verbal signals ( & #33; & #33; & #33; & #33; and #33;) and tone of voice. Due to the special 2D audio model, Du-Jib captures doubts ("ah", "hmm"), sighs, logical tension, sarcasm and laughter. The model took place for the first time in modules of paralinguistic recognition with a score of 48.6 (LISTEN, StressTest, MSPB), before 3.8 Flash (46.9) and Qwen3-Omni (44.6). At the same time, the model not only recognizes emotions, but also organically laughs, pauses, sighs and clears its throat in response instead of reading text tags out loud. The developers introduced a single 32B brain in three versions of Generation Audio. The recommended Tg generates voice tokens tagged by an expressive voice model (NV-SuperBench 0.86/0.80, first audio delay -0.6 s, 16 sessions per 80 GB GPU). The flash version is designed for scale and the automotive industry: the delay of the first audio is reduced to a record 0.31 seconds with the support of 40 simultaneous users on one 80 GB card. The E2E version is a fully experimental end-to-end voice synthesis within a single weighted matrix without a separate speech module. In VoiceBench (82.8) and Big Bench Audio (81.8%), Sophon showed the best result among open models, comparing the closed GPT-4o-mini-Audio from OpenAI. In tests using the full Duplex-Bench v3 tool, the model reached Pass@1 48% with 100% interactive response, beating Grok (43%).

Sophon-Realtime-v1: an open model of full voice dialogue with understanding of sarcasm and laughter The architecture is built on the basis of a 32 billion language model, and implements honest full-duplex dialogue (fully complex) speech-to-skeleton: the system continuously listens to the user during his speech, immediately uploads a copy when he stops and immediately falls silent (barge-in) if a person interrupts. The main technological feature of Sophon is the natural perception of non-verbal signals ( & #33; & #33; & #33; & #33; and #33;) and tone of voice. Due to the special 2D audio model, Du-Jib captures doubts ("ah", "hmm"), sighs, logical tension, sarcasm and laughter. The model took place for the first time in modules of paralinguistic recognition with a score of 48.6 (LISTEN, StressTest, MSPB), before 3.8 Flash (46.9) and Qwen3-Omni (44.6). At the same time, the model not only recognizes emotions, but also organically laughs, pauses, sighs and clears its throat in response instead of reading text tags out loud. The developers introduced a single 32B brain in three versions of Generation Audio. The recommended Tg generates voice tokens tagged by an expressive voice model (NV-SuperBench 0.86/0.80, first audio delay -0.6 s, 16 sessions per 80 GB GPU). The flash version is designed for scale and the automotive industry: the delay of the first audio is reduced to a record 0.31 seconds with the support of 40 simultaneous users on one 80 GB card. The E2E version is a fully experimental end-to-end voice synthesis within a single weighted matrix without a separate speech module. In VoiceBench (82.8) and Big Bench Audio (81.8%), Sophon showed the best result among open models, comparing the closed GPT-4o-mini-Audio from OpenAI. In tests using the full Duplex-Bench v3 tool, the model reached Pass@1 48% with 100% interactive response, beating Grok (43%).