AI NewsOpen sourceAnnouncement
Sophon-Realtime is an open voice model that speaks the OpenAI Realtime protocol
Adventists AI released Sophon-Realtime, an open voice model that answers the same WebSocket protocol as the OpenAI Realtime API.
Image: GitHub
Why it mattersA team paying per minute for voice API calls can test the same client against a model it runs itself. The switch costs a GPU and a pilot, not a code rewrite.
A team on the OpenAI Realtime API keeps the same WebSocket client and swaps the endpoint. A Chinese company called Adventists AI released Sophon-Realtime on 8 November. The model is open-source under Apache 2.0 and answers the same protocol.
The GitHub repository was created on 4 November. It has 419 stars and 35 forks on day 13. The release has three builds at 32 billion parameters and three at 4 billion. Each one listens and speaks at the same time, not in turns. The README calls this full-duplex mode.
What it does
The model picks up on how a user says something, not only what they say. The README writes that the voice laughs, sighs and breathes. Adventists AI publishes a benchmark it calls paralinguistic understanding. The test checks whether a model notices emotion, tone and pauses.
Sophon-Realtime scores 48.6 on that test. Adventists AI says this beats Gemini 3.8 Flash at 46.9 and Qwen3-Omni at 44.6. All three figures are Adventists AI's own.
On VoiceBench, a published voice benchmark, the 32-billion builds score 82.8 percent. Adventists AI says the result is level with GPT-4o-mini-Audio. The company also runs Full-Duplex-Bench v3 for tool calling. It reports Pass@1 of 48 percent, which it says is ahead of Grok at 43 percent. Both numbers are Adventists AI's own.
What you need to run it
Each 32-billion build needs one 80 gigabyte GPU. The 4-billion builds run on 24 to 32 gigabytes of GPU memory. A Mac with Apple silicon and 36 gigabytes of unified memory can run the model offline. The Mac path uses Apple's MLX framework.
The first audio reply arrives in about 0.25 seconds on the E2E build. The Flash build replies in 0.31 seconds. Adventists AI measured both with one user on one H100 GPU. Latency grows when more users share the same GPU. The README does not give a figure for that case.
Voice cloning and languages
The model clones a voice from a reference clip of 3 to 10 seconds. The README lists seven languages ready for conversation. They are Chinese, English, Italian, German, Spanish, French and Portuguese. The Flash build adds Japanese and Russian. Korean is marked as coming soon.
Adventists AI runs a hosted API trial with three free hours on the Flash build. The company lists its weights on Hugging Face and ModelScope under Apache 2.0.
A product team paying per minute for voice calls has one short test to run. Keep the OpenAI Realtime client. Point it at Sophon-Realtime on one 80 gigabyte GPU. Run ten real customer calls through both endpoints with the same prompt. Listen for the parts that matter. Does the model stop when the user interrupts? Does it get the emotion right? Does the first audio arrive fast enough? The vendor numbers above are Adventists AI's own. The ten calls are the test that decides the switch.
Source
adventists-ai/sophon-realtime on GitHub, Apache 2.0.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.