Open full-duplex voice model Sophon-Realtime tops open benchmarks, Apache-2.0 with tone-aware speech
pmttyji · reddit · 2026-11-10
Adventists AI released Sophon-Realtime v1, an Apache-2.0 open full-duplex voice family built around one 32B brain that listens, decides turn-taking and drafts replies. Three builds differ only in how replies become sound: Tag (expressive speech with tags, NVV-SuperBench 0.86 zh / 0.80 en), Flash (0.31s first audio, 40 users per 80GB GPU) and E2E (research preview). It leads open full-duplex models on VoiceBench (82.8) and Big Bench Audio (81.8%), ranks #1 on paralinguistic understanding (48.6, ahead of Gemini 3.8 Flash and Qwen3-Omni), and hits 48% Pass@1 on Full-Duplex-Bench v3 tool calling, beating Grok. Served via vLLM-Omni in one command, free for commercial use.
Related event: Open-Source Full-Duplex Speech Model Sophon-Realtime Released(2 posts)→
More from Models
- pyannote speaker-diarization-3.1 trends on Hugging Face — pyannote · 2026-11-10
- Fable 5.2 draws every pixel via code — no diffusion, no 3D, no render engine — DimitrisPapail · 2026-11-10
- Vibe-coded Google Earth in JavaFX shows Gemini 3.8 Flash is the best free workhorse — AnahataASI · 2026-11-10
- Fable 5.2 after one day: better than Opus at figuring things out, but overthinks simple tasks — flavioAd · 2026-11-10
- Gemini 4 Carbon goes live in Google AI Studio, requires Ultra plan or API key — cedric_chee · 2026-11-10
- Cutting GPU power 42% costs nothing: RTX 3080 Ti runs Qwen 3.5 9B at full speed on 250W — Asthenia5 · 2026-11-10