Moss unveils MOSS-Live to chat and multitask at the same time

MOSS-Live is a full-duplex real-time interaction system by Moss that lets you talk to an AI just like a real phone call, while it works on complex tasks and tool calls in the background.

Beyond voice, it handles live camera feeds, screen sharing, and ambient noise for multimodal understanding. Most importantly, it breaks away from walkie-talkie style turn-taking, allowing smooth, simultaneous listening and speaking.

What it can do

Most voice assistants struggle with conversational rhythm. They cut you off the moment you pause to think, or halt completely if you just utter an agreeable "uh-huh".

MOSS-Live fixes this by parsing semantic intent on the fly. In the FDB v1 benchmark, its pause interruption error rate is just 1.44%, noticeably lower than GPT Realtime 2.1 at 3.83% and GPT-Live-1 at 10.37%. On the τ-Voice test, its resilience to filler words hits 98.17%, while resistance to catchphrases and irrelevant noise reaches 82.16% and 78.90% respectively.

That means you can pause mid-sentence without losing your turn, or murmur in agreement while it keeps talking. But if you actually ask a brand new question, it stops immediately to listen.

It also remembers where you left off after multiple interruptions or side topics, scoring 4.88 out of 5 on FDB v1 post-interruption answer quality. In noisy cars or group chats, its false trigger rate on background voices drops to 2%. It also supports streaming simultaneous interpretation and proactive cues, such as correcting errors during recitation or chiming in during a meeting when specific conditions are met.

How it works: Frontend-backend split

Instead of forcing one model to do everything at once, MOSS-Live uses a smart frontend-backend collaborative design.

The frontend model handles lightweight, real-time sensing. It monitors incoming speech, camera frames, and background audio to decide exactly when to speak, pause, or yield the floor.

The backend model handles heavy reasoning and asynchronous tool calling. If you ask it to plan an itinerary while checking hotel rates, the frontend keeps chatting with you while the backend does the heavy lifting.

A shared coordination layer syncs context and progress. Once the results are ready, MOSS-Live naturally introduces them back into the conversation without making you wait in silence. In testing, its FDB v3 tool call success rate hit 51%, and τ-Voice multi-turn task success reached 57.93%.

Availability and hands-on

Moss has launched both an interactive web demo and API access on its platform, with temporary free API tiers available for testing.

Users can head to the official Moss website, grant microphone and camera permissions, and test real-time video calls or screen sharing right in the browser. Developers can grab an API key through the developer portal to integrate full-duplex conversational audio.

Note that Moss currently provides cloud API and hosted web access only; standalone model weights have not been released for local deployment.

Why it matters

For anyone building real-time AI agents, meeting assistants, in-car voice systems, or live tour guides, MOSS-Live offers a compelling preview of where voice AI is headed.

Its real breakthrough is eliminating the awkward silence that usually happens when an AI pauses to run a tool. Separating conversational flow from background task execution makes live interaction feel fluid and human.

We think Moss made a sharp move by focusing on latency, filler rejection, and interruption recovery first. The main question now is whether they will eventually offer open weights or self-hosted options, which developers will need before putting it into mission-critical enterprise workflows.