MOSS-Live brings voice, vision, and background tasks together in one assistant

Talking to most voice assistants still feels like using a two-way radio. First the person speaks, then the machine. If the user interjects with “uh-huh,” the system may cut off its response; if the user actually interrupts it, the system often loses context and starts over. MOSS-Live is designed to eliminate precisely this interaction pattern: the assistant listens, speaks, and performs tasks simultaneously instead of turning the dialogue into a queue of audio clips.

The OpenMOSS team has released MOSS-Live in beta, including API access. The model accepts not only audio but also visual input from a camera or screen, supports simultaneous interpretation, and can continue a long-running operation in the background while the conversation moves on. This makes it interesting not merely as another voice chatbot, but as a foundation for interfaces in which speech, images, and actions are connected within a single session.

What full duplex changes

Consider a customer-support assistant. A user points a camera at a router and asks where to plug in a cable. The assistant begins explaining, but the person interrupts: “Wait, mine has a different port.” A conventional system either will not hear the comment until it finishes responding or will stop and forget where it left off. MOSS-Live claims more natural behavior: the model distinguishes attention signals from genuine interruptions, accommodates pauses, and resumes the conversation at the appropriate point after the clarification.

For a product, this is not a cosmetic feature. An incorrect interruption can break a voice workflow just as quickly as a frozen button can break a graphical interface. Full duplex could be useful in interpreters, educational apps, customer support, on-screen assistants, and services for people who find it inconvenient to switch constantly between voice and keyboard.

The visual channel broadens the range of use cases. A user can show a faulty component, an app mockup, or a spreadsheet on the screen and then discuss what is visible by voice. Background tasks make it possible, for example, to start analyzing a document while continuing to clarify requirements, without silently waiting for the operation to finish.

The key is a model family, not a single model

OpenMOSS is developing several product lines: MOSS-TTS for speech synthesis and cloning, MOSS-Audio for audio recognition and analysis, MOSS-VL for images and video, and MOSS-Image for image generation. MOVA has also been released separately—a model that generates video with sound; its code and weights are available through the project repository.

This family suggests a possible architecture for a future product: specialized models are connected around a shared interactive interface. A developer does not necessarily have to assign everything to one enormous model. Speech recognition, vision, voice synthesis, and media generation can be separated into individual components whose latency can be measured and which can be replaced independently.

What cannot yet be considered ready

Some of the plans emerged during the team's AMA on Reddit. They include MOSS-Anime, a new image-editing model; MOSS-TTS 2.0 with emotion control; and a compact video-with-sound generator for consumer graphics cards. These are useful indicators of the development roadmap, but they are not finished products. Until the weights, license, memory requirements, and independent tests are published, it would be risky to build a roadmap around these promises.

MOSS-Live itself should also be approached as a prototype. For a voice product, impressive demonstrations are not enough: response latency, interruption-detection quality, resilience to noise, per-minute cost, language coverage, and the rules governing camera and microphone audio processing all matter.

The practical next step is straightforward: choose one short workflow—for example, voice guidance while working on screen—and test it with real interruptions, pauses, and background tasks. If the API consistently preserves context without introducing noticeable latency, MOSS-Live could become a useful layer in a multimodal product. The rest of the announced lineup should reasonably be evaluated only after releases and reproducible tests.