MOSS-Live Released: It Can Listen, Speak, and Get Things Done at the Same Time, but Don’t Rush to Call It Jarvis Yet

MOSS-Live has launched. MOSS (Mosi Intelligence) defines it as a full-duplex real-time interaction system: it keeps listening while speaking, and background tasks do not block the conversation. The official product page and demo entry are at the end of the article.

Most voice assistants are like walkie-talkies: you finish speaking, then they speak. MOSS-Live’s selling point is more like a phone call, with listening and speaking happening in parallel. The official materials list several scenarios: if you pause to think, it waits; if you are just acknowledging, it keeps talking; if you genuinely interrupt, it stops immediately; if people nearby are chatting, it does not cut in. After being interrupted, it remembers the exact word it actually reached and continues from that point, without assuming you heard content it never said.

It also supports video calls and screen sharing, and can recognize ambient sounds and voice characteristics. While you speak, it translates in real time. Background tasks can check the weather and stock prices, and can also run longer research and planning; when results are ready, it brings them back into the conversation at an appropriate time. Rules can be set, such as when to correct errors, count, or remind.

The official architecture is a front-end/back-end collaboration. The front-end model handles real-time listening, speaking, and seeing, while the back-end model asynchronously performs reasoning and tool calls. The two share memory, context, and task state.

Officially, it leads on six metrics among the compared models. Specific numbers: FDB v1 takeover rate down 1.44%; FDB v1.5 response rate to bystander speech down 2.00%, recovery rate up 68.00%; τ-Voice backchannel selectivity up 98.17%, filler-word selectivity up 82.16%, non-directed speech selectivity up 78.90%. These are official self-evaluations, not third-party reproductions.

Among the scenario demos, the Shanghai sightseeing segment is relatively concrete. It plans a route, recognizes the Peace Hotel via real-time video, explains its history, and looks up room rates and dining on demand. When the tourist asks where it left off, it can continue from there.

In the meeting scenario, when a colleague adds context, it adjusts its research, reports back once results are ready, handles interruptions and acknowledgments, and finally revisits earlier topics to give a summary and action items.

In the road-trip scenario, it updates the itinerary, checks hotels in the background, answers questions about ferries and parking, returns to the route description that was interrupted, and can keep speaking over navigation audio.

The comments section focuses on more practical concerns. X user nynach asks, isn’t it going to be open source?; kshabana directly asks for open source. LastResort asks about cost per minute. Trying my best :-) questions that it cannot be hosted locally, saying they do not want their child’s conversations and personal thoughts uploaded. AI Tools Daily says they have tested several voice agents, the most common failure is interruption handling, and they most want to test that first in a noisy room. Ruslan asks whether, when a background task finishes while the user is speaking, it will wait for a natural pause before speaking; even a fast assistant can seem rude.

One developer’s feedback is worth reading. Sebastian Sosa says the most painful pitfall when doing things while chatting is when the voice says “I’m doing it” but does not hand the task off to a background worker; it happened twice in a row in the logs. He used a small model to read each moment and make the handoff, reducing “go back” from 45 seconds to 3 seconds.

The direction MOSS-Live demonstrates is clear: voice interaction is moving from turn-based to continuously present. But open source, cost, local deployment, and the reliability of task handoff are still unanswered in the release. To try it, go to the MOSS-Live demo entry or API Playground. Don’t rush to call it Jarvis yet; first test it in a noisy room and around background-task completion timing.