Skip to main content

Meta's Muse Spark 1.1: Smarter Multi-Agent AI That Can Fix Its Own Code

Meta just dropped a new flagship AI model, Muse Spark 1.1, and it's all about making multi-agent automation smoother and smarter. Think of it as the conductor of an AI orchestra—it can orchestrate a main agent and several sub-agents to work together on complex tasks, adjusting plans on the fly when things change.

What Makes Muse Spark 1.1 Special?

Multi-agent workflows aren't new, but they've always had a headache: when agents generate tons of data, they often hit a context limit and have to ditch some info, which hurts quality. Muse Spark 1.1 tackles this with a clever context compression trick. It squeezes the data while keeping the important bits, so it can refer back to earlier work without losing the plot. The model's context window now holds up to a million tokens—plenty of room for even the messiest projects.

Coding Like a Pro

This model really shines in coding tasks. In an internal test, Meta engineers asked it to build a chat app from a simple prompt. Muse Spark 1.1 didn't just write code—it automatically took screenshots of the app's interface, spotted bugs, and hunted down the exact code snippets causing trouble, then fixed them. Talk about a self-sufficient developer!

On the Vibe Code Bench v1.1, a benchmark for AI programming, Muse Spark 1.1 scored 72.2—more than 50 points higher than Meta's previous flagship. It also jumped nearly 18% on the SWE-Atlas Codebase QnA test. That's not just an improvement; it's a leap.

Beyond Code: Real-World Tasks

But it's not just about coding. Muse Spark 1.1 can handle other multi-step jobs, like generating product descriptions from videos or placing restaurant orders for you. It's like having a super-efficient assistant that never forgets what you asked for.

Availability and What's Next

Developers can already try Muse Spark 1.1 through the Meta Model API. And Meta isn't stopping here—they're planning to boost data center capacity to 14 megawatts next year and are expected to launch their own AI chip, codenamed Iris. The future looks busy for Meta's AI ambitions.

Image

Key Points

  • Context compression: Handles up to 1 million tokens without losing important data.
  • Multi-agent orchestration: Main agent plans, sub-agents execute, and the model adapts in real time.
  • Coding prowess: Scored 72.2 on Vibe Code Bench, a huge jump from previous models.
  • Available now: Public preview via Meta AI chatbot and API.
  • Future plans: Data center expansion and a custom AI chip called Iris.