AI Briefingadventists-ai/Sophon-Realtime-v1 · Hugging Face
·2026.11.10 23:20
Key point
The Apache-2.0 licensed model family offers three builds, with the Flash variant achieving 0.31 s first audio latency and supporting 40 concurrent users per 80 GB GPU.
Details
- Adventists AI released Sophon-Realtime v1, an open full-duplex voice model family based on a shared 32B parameter brain that handles listening, turn-taking, and reply generation.
- The model achieves a VoiceBench score of 82.8 and Big Bench Audio score of 81.8%, ranking #1 among open full-duplex models and matching GPT-4o-mini-Audio on VoiceBench.
- Three builds are available: Tag (recommended for expressive voice), Flash (optimized for speed and scale), and E2E (research preview for end-to-end speech).
- The Flash build delivers first audio in 0.31 s and supports 40 concurrent users per 80 GB GPU, while the Tag build supports 16 users per GPU and achieves NVV-SuperBench scores of 0.86 (Chinese) and 0.80 (English).
- Code and weights are released under the Apache-2.0 license, free for commercial use, and served via vLLM-Omni with compatibility for the OpenAI GPT-Live interface.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
GitHub Next releases Ace, a local collaborative coding environment for humans and AI agents
Gnomon: Git-like version control for Postgres data built for AI agents