Skip to main content

Meta's Muse Spark 1.1: Smarter Multi-Agent AI That Keeps Its Cool Under Pressure

Meta has officially launched its new flagship large language model, Muse Spark 1.1, designed to optimize multi-agent automation workflows. The model is now available for public preview through the Meta AI chatbot service and API interface.

Image

Multi-agent automation workflows typically consist of a main agent and multiple sub-agents, where the main agent is responsible for creating the task execution plan, while the sub-agents perform specific operations based on instructions. At the beginning of a new project, the main agent generates the project plan. Meta states that Muse Spark 1.1 can detect and respond to changes during execution in a timely manner, thus adjusting the project plan.

Since multi-agent automation tasks often involve many steps, agents generate a large amount of data during execution. If this data exceeds the model's context limit, some information must be discarded, which usually affects output quality. Muse Spark 1.1 solves this challenge with a context compression mechanism, compressing generated data in a way that preserves the most important details. This allows it to retrieve information from earlier work when needed, effectively passing data between different sub-tasks. The model's context window has reached one million tokens.

The context compression and multi-agent capabilities of the new model make it perform well in coding tasks. In an internal test, Meta engineers asked it to generate a chat application based on a prompt, and Muse Spark 1.1 was able to automatically capture screenshots of the program interface, identify technical issues, and find the code snippets causing the problem for repair.

Muse Spark 1.1 scored 72.2 on the AI programming benchmark test called Vibe Code Bench v1.1, far exceeding Meta's previous flagship model by more than 50 points. At the same time, its score also increased by nearly 18% in another test, SWE-Atlas Codebase QnA.

This model can complete other multi-step tasks while generating code, such as generating e-commerce product descriptions based on product videos and placing restaurant orders on behalf of users. Developers can easily access Muse Spark 1.1 through the Meta Model API. According to reports, Meta plans to increase data center capacity to 14 megawatts next year and is expected to launch a self-developed AI chip called Iris.

Key Points

  • Multi-Agent Mastery: Muse Spark 1.1 handles complex workflows with a main agent and sub-agents, adapting plans in real time.
  • Context Compression: A new mechanism compresses data to retain crucial details, enabling a 1 million token context window.
  • Coding Prowess: Scored 72.2 on Vibe Code Bench v1.1, a 50+ point jump over its predecessor, and improved 18% on SWE-Atlas Codebase QnA.
  • Real-World Tasks: Beyond coding, it can generate product descriptions from videos and even place restaurant orders.
  • Availability: Now in public preview via Meta AI chatbot and API; Meta also plans to expand data centers and launch its Iris AI chip.