Introduction

OpenMOSS / MindSmart (模思智能) announced MOSS-Live on 2026-11-02 — a full-duplex realtime voice interaction system meant to listen and speak at the same time, not walkie-talkie turn-taking. The vendor X post is the zero-day signal; the product page and hosted playground/API are live today.

This is a closed / API voice-agent ship for makers who want interruptible calls with tools and vision — not an open-weights TTS drop.

What Shipped

  • Vendor / product: MOSS-Live — OpenMOSS / MindSmart
  • Modality: realtime voice agent (full-duplex STT+TTS interaction + tools / vision)
  • Open vs closed: closed / API (hosted demos + Playground + API; no open generative weights in this drop)
  • Runnable today: web call demos (Luna / Leo) on moss.cn; Playground + API on platform.moss.cn as MOSS-Live-1.0 beta with limited-time free access per the vendor
  • Concrete fact: in the vendor’s compiled Full-Duplex Bench (FDB) comparison they highlight pause barge-in 1.44%, unrelated false-accept 2%, and FDB v3 tool-call success 51% (tied best in their table); foreground/background split keeps long tools running while the conversation continues

Primary announce:

What Full-Duplex Means Here

  • Semantic interruption: keeps going through backchannels like “mm-hmm,” but stops for a real new question.
  • Multi-speaker / ignore: tries not to answer navigation prompts or side talk aimed at someone else.
  • Multi-turn memory: resumes an earlier thread after you digress.
  • Audiovisual: video calls and screen share feed the same conversation loop.
  • Async tools: research / booking-style jobs run in the background; results re-enter when ready.
  • Simultaneous interpretation is listed among the product capabilities on the model page.

Treat the FDB / τ-Voice-style numbers as vendor-compiled until independent benches catch up.

Try It

  1. Open MOSS-Live product and start a Luna or Leo call demo.
  2. For API work, use platform.moss.cn — MOSS-Live-1.0 beta, limited free window per @MOSS_Global.

No open weights to pip install this cycle; the runnable surface is the hosted stack.

Limits Worth Knowing

  • Closed model — you cannot self-host the generative stack from this announce.
  • Beta API / free tier can change; pin your own latency and interrupt tests.
  • Bench tables on the marketing page are not a substitute for your domain eval (accents, noise, tool schemas).

Original Source

Conclusion

If you need a hosted full-duplex voice agent you can dial today — with barge-in, optional video/screen context, and background tools — MOSS-Live is the November 2 closed stack to try while open on-device TTS (Linnet, Klein, Kuyil…) covers a different maker job.

—Anabel ♡