Building the Next Gen
Voice Agent OS
Giving AI a true personality, memory, and perception. More than a tool—a digital life that understands and accompanies you.
Sophon-Realtime: hears your tone.
Speaks with one.
🆕 v1.1 (Nov 11): zero-shot voice cloning. A few seconds of audio, and it talks to you in that voice in real time.
Sophon-Realtime is our open full-duplex voice model family. It listens while you talk, hears how you say things, and answers with a voice that laughs, sighs and breathes when the moment calls for it. Tag and Flash run on one 80 GB GPU with vLLM-Omni. Code and weights are Apache-2.0, free for commercial use.
v1.1 demo: 11 s of our founder's voice, then 9 s of a colleague's, swapped in without a restart. Real-time responses, not sped up.
First audio with the Flash build. It drafts the reply while you are still talking.
VoiceBench, #1 among open full-duplex models (Big Bench Audio 81.8%)
Paralinguistic understanding, #1 overall, ahead of Gemini 3.8 Flash (46.9)
per 80 GB GPU with Flash (16 with Tag). One GPU, one command.
pip install "sophon-realtime[trt]" && sophon-realtime serve
Benchmarks are self-run unless noted; methods and every chart are in the GitHub README. Latency and concurrency measured on one H100 80 GB. Voice cloning: only clone your own voice or voices you have permission to use; do not use it to impersonate anyone or to deceive.
A System One model
for speech.
🆕 Oct 22: now hears Japanese. Japanese error 33% → 12%, Korean 32% → 7.6%, Vietnamese 42% → 15%
Jev made decision models the idea of the year: no text generation, just typed, probabilistic answers, fast and cheap. But a text decision model only sees the transcript, and the transcript is where tone, stress and hesitation disappear. DuplexJev decides straight from raw audio: ten decisions in one forward pass, and ahead of six models from Google, OpenAI and Alibaba on public paralinguistic benchmarks.
Paralinguistic understanding: three public benchmarks combined
LISTEN (EACL 2026) ×4 + StressTest + MSPB, 1,035 questions, none from our training corpora, same for every model, higher is better
⚠ Both realtime flagships trail their own offline models: the skill a voice agent needs most is where they are weakest.
Ten decisions in one forward pass; transcribe-then-generate takes 1,978 ms (21× slower)
per million decisions with the 4B model; a new decision is one line of text
Prosody (MSPB: statement or question, rise and fall), highest of seven models; next best Qwen3-Omni, Gemini 3.8 Flash 67.2
Which word is stressed (Mandarin), next best Gemini 3.8 Flash; English 96.0 vs 88.5
Same questions; closed models asked once, first option letter taken; DuplexJev reads option probabilities. Gemini 3.8 Flash leads StressTest alone (89.0 vs 86.5). Stress and emotion sets come from corpora in our training data (held-out speakers or sentences). DuplexJev is an independent project by Adventists AI and is not affiliated with TypeSafe (Jev).
Speed is the First Sense of Intelligence
In human conversation, latency over 200ms creates a sense of disconnection. Traditional voice assistants often take 2-3 seconds to respond. We've engineered Adventists AI to respond at human speed.
500ms Ultra-Low Latency
Using our cutting-edge Dynamic Sparse Activation architecture, parameter activation flows smoothly from 3B to 275B based on thought depth. This achieves a Time-to-First-Token (TTFT) of just 500ms—faster than average human reaction time.
Think and Remember Like a Human
Thinker-Talker Parallel Reasoning
- Parallel Strategy: Up to 16 parallel calculation paths running even while speaking.
- Intuition & Logic: Speaking with intuition while correcting with logic in real-time.
- Model Interop: Emotion, logic, and memory models communicate instantly to decide the conversation flow.
Lifetime Memory Palace
Memory isn't just storage; it's about surprise and connection.
We use knowledge graphs, emotional vectors, and forgetting curves to simulate human memory, allowing AI to recall shared moments and build long-term relationships.
Deep Persona & Infinite Knowledge
Define over 40 persona dimensions—from speech style to subconscious values.
Attach "Skill Books" via our dual-layer recall mechanism (10ms latency) to give your agent instant professional expertise.
10-Second Voice Cloning
Break the boundary between virtual and reality. Clone any voice with just a 10-second audio clip.
Our engine captures not just the timbre, but the emotional cadence and breathing rhythm. Instantly attach cloned voices to any NPC.
MCP Ecosystem: The Neural Hub
Connecting the digital and physical worlds through the Model Context Protocol.
First-Party MCP
Includes Weather, News Briefs, Voice Accounting, Daily Fortune, Web Search, and more.
Edge-Side MCP
Supports up to 20-step chain calls and ReAct complex control logic for IoT, lights, volume, and more.
Third-Party MCP
One-click mount thousands of community tools in under 10 seconds.
Hear You, See You, Care for You
Zero-Latency Vision
Unlike the industry standard 10s delay, our streaming video understanding captures expressions, gestures, and environment in real-time.
Identity Recognition
Voiceprint + Face Recognition. The AI knows who you are naturally during conversation, without awkward logins.
AlwaysOn & Proactive
Breaking the "turn-taking" mold. Your AI companion can initiate conversation, greet you when you arrive, and comfort you when you're down.
Competitive Enterprise Solutions
1/1000
The cost of our Realtime API is 1/1000th of OpenAI's pricing.
300+
Hardware products powered by our ecosystem.
Zero
Startup cost with our hardware control board support program.
Powered by Adventists AI
Dipal D1
World's first curved-screen 3D AI character pod. $2.5M crowdfunded on Kickstarter, founded by Vega.
View Details →
Doumaomao DiMi
AI intelligent companion with 10 years of experience, 50+ patents, 200+ team members. Led by CEO Chen Yalang.
View Details →
WaterDrop Bot
45g ultra-lightweight with 4G independent connectivity. Truly portable AI companion by CEO Mao Xin.
View Details →
Xiaoxiaole Series
Museum AI companion series. 23 partner museums, 40M+ visitors served. Bringing cultural relics to life.
View Details →One System, Infinite Possibilities
From simultaneous interpretation to creative image generation, from meeting minutes to smart translation. Every mode is carefully crafted and ready to use.
Seamless Cross-Language Communication
Real-time multilingual interpretation across Chinese, English, Japanese, Korean and more. Millisecond-level latency for international conferences, cross-border business, or daily conversations.
From Imagination to Image, in One Sentence
Generate or edit images directly within the conversation — no app switching needed. Describe your vision and AI delivers visual creations instantly.
Live Transcription
Real-time speech-to-text for multi-person meetings with precise speaker identification
Smart Summary
AI extracts core topics, key decisions, and discussion highlights automatically
Action Items
Automatically identifies to-dos with assigned owners and deadlines
Multi-Format Export
One-click export to Markdown, PDF, or sync to collaboration tools
Say Goodbye to Manual Note-Taking
AI listens to your entire meeting in real-time, automatically distills key decisions, action items, and to-dos into structured meeting minutes.
Sophon System
Not just conversation — AI that actually does things.
Say one command to your Adventists device. Mac handles the rest.
Ready to Start?
Whether you're integrating APIs or building hardware, we're ready.
Australia: +61 411980594
China: +86 18601711715
Suite 10/1 Main Street, Box Hill, Vic, 3128
Lehui Innovation Center, Bantian, Longgang, Shenzhen
Guangqi Park, Xuhui District, Shanghai