Sophon-Realtime-4B-Flash
No GPU? Try it through the hosted API: sign up and get 3 hours of free voice chat (OpenAI Realtime-compatible, docs).
Version 1.4.0 (2026-11-16) · changelog · GitHub
The 4B member of the Sophon-Realtime family: the same full-duplex design as the 32B builds, on one 24 GB GPU (uses ~21 GB). It listens while you talk,
decides from your voice when you have finished, writes a reply with sound tags ([laughter] [cough] [sigh] [breath] <strong>…</strong>)
and speaks while it writes; a lighter, faster speech model speaks the tagged reply (one forward pass per 80 ms frame). Served as one vLLM-Omni model, with /v1/realtime and /v1/chat/completions.
Sophon-Realtime 4B: three builds, one 4B brain
| Tag | Flash | E2E | |
|---|---|---|---|
| How it speaks | an expressive speech model (tag-tts) performs the sound tags | a lighter, faster speech model speaks the tagged reply (one forward pass per 80 ms frame) | a compact talker (about 350M parameters) turns the tagged reply straight into voice vectors; a decoder turns them into sound |
| Runs on | one 32 GB GPU (uses ~27 GB) | one 24 GB GPU (uses ~21 GB) | one 32 GB GPU (uses ~31 GB) |
| Full-Duplex-Bench v1.0 (4-task avg) | 94.3 | 87.7 | 92.3 |
| First audio, one user (RTX PRO 6000) | 0.4 s | 0.21 s | 0.31 s |
| Weights | HF · ModelScope | HF · ModelScope | HF · ModelScope |
The 4B brain is DuplexJev-4B-Para, our open 4B speech model, with a tag-writing adapter merged in. Speech only: the 4B brain has no vision. For the strongest answers, use the 32B builds (Tag · Flash · E2E) or the hosted API.
Benchmarks
| 4B | 32B (published) | |
|---|---|---|
| VoiceBench overall | 62.4 | 82.8 |
| Big Bench Audio (spoken answer, transcribed) | 44.3% | 81.8% |
| Full-Duplex-Bench v1.0, Flash build (avg of 2 pause tasks, turn-taking, interruption) | 87.7 | 94.0 |
| Sound-tag rate on held-out conversations (reference: 0.36) | 0.36 | — |
What 4B is for. The 4B brain keeps the conversation skills of the family: listening while you talk, deciding when you have finished, back-channels, barge-in and sound tags (Full-Duplex-Bench close to the 32B builds, and the same tag rate as the reference data). It knows and reasons much less than the 32B brain (VoiceBench 62.4 vs 82.8, Big Bench Audio 44.3% vs 81.8%), so it can state wrong facts with confidence. Use it for local demos, companions and characters, and research on turn-taking; use the 32B builds or the hosted API when answers must be right. Among open full-duplex models, its VoiceBench score is above Freeze-Omni (55.2) and Moshi (29.5) and below MiniCPM-o (71.2).
Measured 2026-11-16 on one RTX PRO 6000 with each build held to the memory of the card it is configured for. VoiceBench: official evaluators, gpt-4o-mini judge, AlpacaEval on the standard 199-question subset, overall = mean of 9 tasks. Big Bench Audio: 1,000 questions, GPT-4.1 judges the transcript of the spoken answer. Full-Duplex-Bench v1.0: average of candor and synthetic pause handling, turn-taking and user interruption (backchannel is reported but not averaged), the same scripts as for the 32B release. First audio: one user, client_example.py, median of 6. The 32B column repeats the numbers published for the 32B builds.
Quick start
git clone https://github.com/adventists-ai/sophon-realtime && cd sophon-realtime/flash
bash install.sh # once: Python 3.12 venv, vLLM 0.30, vLLM-Omni 0.30, plugin
modelscope download --model AdventistsAI/Sophon-Realtime-4B-Flash --local_dir ./model # or: hf download adventists-ai/Sophon-Realtime-4B-Flash --local-dir ./model
GPU=0 PORT=8000 bash start.sh # sees the 4B model and picks its 24 GB config
python client_example.py --url http://127.0.0.1:8000
Full instructions, the realtime protocol, zero-shot voice cloning and Live mode: flash/README.md.
Layout
config.json · thinker/ (4B, bf16), talker/, voices/
License: Apache-2.0; see NOTICE for the sources of the bundled weights. Only clone your own voice or a voice you have permission to use.
- Downloads last month
- 106