MOSS-Live explores an AI assistant that can listen while it speaks

OpenMOSS has described MOSS-Live, a real-time assistant designed to listen and speak at the same time, respond to interruptions, process live visual input and continue background tasks while a conversation carries on. The project’s November 7 overview describes a hosted playground and realtime API, not a downloadable local model release.

Two jobs at once

The team describes a foreground interaction model that handles listening, turn-taking and speech alongside a background model for reasoning and tools. A coordination layer shares context and task state. In the team’s travel demonstration, for example, the assistant can discuss a follow-up while a lookup continues. That is a design description and demonstration, not proof that it will reliably understand every interruption or safely carry out every task.

The distinction matters because conventional voice interfaces often take turns: record a user, stop listening, generate an answer and then resume. A system that can listen throughout needs to decide whether a brief acknowledgement is an interruption, and it must protect private audio and video while doing so. The publisher describes selective attention; we have not evaluated it ourselves.

What is available

The public project repository currently provides documentation, demonstrations and links to a playground and API. Its presence does not establish that the full model weights or implementation are open source. Check the provider’s current access, pricing and data-handling terms before sending microphone or camera input.

OpenMOSS reports results on several speech and interaction benchmarks, but those are publisher evaluations, not an independent test. In our view, the interesting engineering question is less whether a demonstration feels seamless for one exchange and more whether it remains accurate, responsive and predictable through interruptions, long tasks and changing permissions.