When AI needs to converse with you in real time while also looking up information, using tools, and completing complex tasks, a new challenge emerges.
Real-time conversation requires AI to understand and respond quickly, while complex tasks often involve multiple reasoning steps, information retrieval, and tool calls that take longer to process. How to complete these tasks while maintaining real-time interaction becomes a challenge that system architecture must address.
To solve this, MOSS-Live uses a foreground–background architecture that assigns real-time interaction and complex task execution to different models. This allows AI to complete complex tasks asynchronously while keeping the conversation going.
What Is a Foreground–Background Architecture?
MOSS-Live’s foreground–background architecture consists of three main components: a foreground interaction model, a background model, and a coordination mechanism.
The foreground handles real-time perception and conversation, while the background executes complex tasks. The two stay coordinated through memory, context, and task status.
Foreground Interaction Model: Real-Time Perception and Conversation
In full-duplex mode, the foreground continues listening while speaking. It uses semantic cues, rather than simple changes in volume, to determine whether the user is interrupting or has finished their turn. In settings with multiple people, it identifies the person it is interacting with and tracks the progress of its own speech output in real time. It also continuously receives speech, visual input, and environmental sounds, adapting the interaction as new input arrives.
Background Model: Complex Reasoning and Task Execution
When a user requests a task that requires multiple steps or complex processing, the background model handles work such as tool use, information retrieval, and extended research. Tasks run asynchronously in the background without blocking the foreground’s real-time conversation.
Coordination Mechanism: Keeping Both Models Aligned Through Memory, Context, and Task Status
The foreground and background share conversation context and memory. Task status updates keep both sides aligned on task dispatch, execution progress, and results. The foreground can check the status of background tasks at any time and present the results to the user at an appropriate moment once a task is complete.

How Foreground–Background Coordination Works
The key to foreground–background coordination is running real-time interaction and complex task execution in parallel. When a user requests a task, the foreground interprets the request and passes it to the background. The background executes the task asynchronously, while the foreground continues interacting with the user in real time.
For example, a user might say: “Help me research the latest developments in the AI industry.”
First, the foreground interaction model understands the request and sends the task, along with the relevant context, to the background model. The background then begins retrieving information and performing complex reasoning. Meanwhile, the foreground remains active, continuing to listen and respond to the user.
Throughout task execution, the coordination mechanism keeps the foreground and background in sync. Execution progress and task status are continuously shared with the foreground. While maintaining real-time conversation, the foreground can also adapt the interaction based on new speech, visual input, and environmental sounds.

Once the background completes the task, it returns the results to the foreground, which presents them to the user at an appropriate moment. Throughout this process, the background handles the complex task while the foreground maintains real-time conversation. Together, they ensure that task execution no longer interrupts the conversation.
Key Benefits of the Foreground–Background Architecture
By handling real-time interaction and complex task execution separately, MOSS-Live can maintain natural conversation while also tackling complex tasks.
-
Uninterrupted Real-Time Interaction: While the background executes complex tasks, the foreground continues listening and responding. Users can keep asking questions and providing additional information without waiting for the task to finish, maintaining a natural conversation in real time.
-
Parallel Execution of Complex Tasks: Through this division of responsibilities, tasks such as deep reasoning, tool use, information retrieval, and extended research can run asynchronously in the background. This allows AI to handle more complex work while continuing to converse with the user.
-
Continuous Context Synchronization: The foreground and background stay aligned through memory, context, and task status management. The foreground can track task progress and share results at an appropriate moment once they are ready, allowing task execution and real-time conversation to flow naturally together.
Through foreground–background coordination, MOSS-Live brings real-time interaction and complex task execution together, enabling more capable task handling while preserving natural conversation.
Try MOSS-Live Now
MOSS-Live is now available through an online demo and API access.
You can chat with MOSS-Live on the official website or integrate it into your own products and applications through the API.
-
Try MOSS-Live Online: https://platform.moss.cn/app/playground/
-
Real-Time Interaction API Documentation: https://platform.moss.cn/docs/scenarios/moss-live/
-
Project Overview and Full Evaluation: https://github.com/OpenMOSS/MOSS-Live