# Is Sophon-Realtime v1 SOTA?

Pending benchmarks. Announced Nov 8, 2026; no independent results yet.

As of Nov 22, 2026, Sophon-Realtime v1 is pending benchmarks on isitsota: announced Nov 8, 2026, with no independent results on the boards we track yet.

Adventists AI's open full-duplex speech-to-speech family. Tag, Flash and an E2E research preview share one 32B thinker that listens continuously, decides turn-taking and writes the reply; they differ in how the reply becomes sound. Apache-2.0 code and weights are public on Hugging Face and ModelScope for self-hosting with vLLM-Omni (Tag/Flash on one 80 GB GPU) or the lab's E2E engine. A ModelScope Studio demo exists; no separate public hosted API is confirmed.

Canonical: https://isitsota.com/m/sophon-realtime
Last updated: 2026-11-09
Tracking since: 2026-11-09

## Announced facts

| Fact | Value |
| --- | --- |
| Lab | Adventists AI (GitHub/HF org adventists-ai; ModelScope AdventistsAI). Contributor accounts include BKtensor / @BKJinAI |
| Announced | Nov 8, 2026 |
| API | Not yet |
| Open weights | Published |
| Parameters | 32B thinker (shared); Tag/Flash/E2E speaking stacks differ |
| Role | Open full-duplex voice model: continuous listening, barge-in, back-channels, and spoken replies with optional non-verbal sound tags |
| Builds | Tag (recommended expressive speech model; ~0.6 s first audio), Flash (lighter/faster speech model; 0.31 s first audio), E2E preview (talker inside the 32B model; ~1 s; two 80 GB GPUs) |
| Lineage | Thinker is DuplexJev-32B-Para v1.5 with tag writing merged in (FP8, vision tower retained). DuplexJev's own card lists Qwen3-VL-32B-Instruct language-model weights plus a Qwen3-ASR-0.6B encoder — not a plain Qwen3-32B text checkpoint |
| License | Apache-2.0 for Sophon-Realtime code and weights (GitHub LICENSE, HF cards, PyPI). The standalone DuplexJev-32B-Para model card is still tagged CC-BY-NC-4.0 — licensing discrepancy versus this Sophon redistribution |
| Access | Public ungated HF/ModelScope weights (Tag/Flash/E2E); GitHub release v1.0; pip package sophon-realtime (Tag serve CLI). ModelScope Studio demo linked from the README. Self-host only for production serving; no separate public hosted API confirmed |
| Serving | Tag/Flash: one 80 GB GPU via vLLM-Omni (lab recommends up to 16 / 40 concurrent users). E2E: own engine on two 80 GB GPUs. Do not run third-party installers from this tracking workflow |
| APIs when self-hosted | ws://HOST:PORT/v1/realtime (OpenAI-Realtime-style full duplex) and POST /v1/chat/completions (audio in, text+audio out); optional Live mode compatible with OpenAI GPT-Live clients |
| Limitations (lab) | Tag card: [sigh] and laughter spans remain weak versus how often the thinker writes them; Mandarin slightly less clear than English with the distilled flow; very long replies (over about a minute of speech in one turn) exceed voice training |
| Related | Family page covers Tag/Flash/E2E. Keep separate from DuplexJev hearing models, tag-tts, and Qwen3-VL-32B-Instruct |
| Date evidence | GitHub release v1.0 published 2026-11-08T05:48:49Z (12:48 AM ET) with Tag/Flash weights linked; HF Tag/Flash family card commits the same morning. Oct 29 HF createdAt and Nov 7 weight-upload commits are not used without proof the repo was public then; the repo's first commit was a teaser saying it opens Nov 9 |

Tags marked unsure are not stated on the lab's own pages.

## What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

- [x] Publicly announced with a source
- [x] Usable by the public (API or weights)
- [ ] Listed on an independent board we track
- [ ] Enough results for a rank in a category
- [ ] Holds across two snapshots (crowns only)

## What the lab claims

Self-reported. Shown for context; not used in any verdict.

- Lab claim: self-reported VoiceBench 82.8 and Big Bench Audio 81.8% for the shared 32B brain (README / family card; self-run unless noted).
- Lab claim: Flash first audio 0.31 s; Tag about 0.6 s (README table; one-user figures).
- Lab claim: Tag NVV-SuperBench 0.86 Chinese / 0.80 English; Full-Duplex-Bench v3 tool-calling Pass@1 48% with 100% response rate (README; self-run).

## Sources

- REPO: [Adventists AI Sophon-Realtime GitHub (Apache-2.0)](https://github.com/adventists-ai/sophon-realtime)
- REPO: [Sophon-Realtime v1.0 GitHub release](https://github.com/adventists-ai/sophon-realtime/releases/tag/v1.0) (2026-11-08)
- REPO: [Sophon-Realtime-32B-Tag model card and weights](https://huggingface.co/adventists-ai/Sophon-Realtime-32B-Tag)
- REPO: [Sophon-Realtime-32B-Flash model card and weights](https://huggingface.co/adventists-ai/Sophon-Realtime-32B-Flash)
- REPO: [Sophon-Realtime v1 family card](https://huggingface.co/adventists-ai/Sophon-Realtime-v1)
- REPO: [DuplexJev-32B-Para card (lineage and CC-BY-NC-4.0 tag)](https://huggingface.co/adventists-ai/DuplexJev-32B-Para)
- LAB: [Author announcement on X](https://x.com/BKJinAI/status/2119673787700699235) (2026-11-09)

Tracking continues until an independent board lists it.

## Related

- Lab: [Adventists AI](https://isitsota.com/labs/adventists-ai.md)
- All pending releases: https://isitsota.com/releases.md
- JSON: https://api.isitsota.com/v1/models/sophon-realtime
