100Combines context management with parallel sub-agents to connect long hours of work to external tools and screen operations.AIIt is available at Muse Spark.1.1directed to developers.APIAlso available in public preview.
Offer an agent model for developers with API
The multi-modal inference model for agent-type tasks, Muse Spark 1.1 connects the output of planning and inference to external tools and operations on the computer. The developer’s Meta released it as a model enhancing tool use, computer operation, coding, and multi-modal understanding.
For developers, the public preview of the Meta Model API, which can be connected to the Muses Spark 1.1, has begun.
Incorporate new tools into the plan and assign them in parallel.
The use includes native tools, MCP servers, and custom skills.Muse Spark 1.1 is said to be zero-shot and integrated into work, according to Meta.
When running, the main agent gathers the necessary context, makes a plan, and delegates the processing to a parallel subagent. Each subagent handles according to the assigned role and available tools, and returns to the main agent if a superior judgment is needed.
100Million tokens to retain the information needed for subsequent work
The 100 million token context window supports long hours of work.Muse Spark 1.1, according to Meta, actively manages this area and retrieves the necessary information from previous work.
If the amount of information increases, compress the context to leave the important steps necessary for subsequent work.This management is also used for operations across multiple apps, and even if the information and requirements change during a long session, it will maintain the context and adapt to updated conditions, Meta explains.
Screenshots and direct operations.
When running a computer, the Muses Spark 1.1 is not designed to process all the procedures on the screen at once. It is trained to create scripts with automated processing, and to use an interface with direct and easy processing.
The method of operation selected brings together and generates multiple actions needed for each stage.In the performance of the dinner arrangement shown by Meta, the model detected the situation changes that occurred during the order and updated the order content without user intervention.
Connect the errors on the screen to code correction and verification.
In coding, Meta explains how to address complex code, diagnosis and correction of defects, implementation of new functions, and large-scale code migration.
In the performance shown by Meta on OpenCode, Muses Spark 1.1 built a web app that automatically captured the screen and identified the user’s visible defects.
Deliver the visual information of the video to the browser.
Multimodal processing examines the visual and audio information to be operated and retains the details until the end of the process.Muse Spark 1.1 uses the information retained as input when operating the computer on behalf of the user.
In the product performance to Facebook Marketplace shown by Meta, they extracted the available product photos from the videos taken on the smartphone and inferred about the products shown, handing the output over to the browser and creating the product instead of the user.