Meta released Muse Spark 1.1, a multimodal reasoning model designed for agentic tasks, marking a significant upgrade from its predecessor Muse Spark.

The model delivers major improvements in tool use, computer automation, coding, and multimodal understanding. Meta Superintelligence Labs built Muse Spark 1.1 to handle complex workflows across multiple applications while maintaining context over extended sessions.

Enhanced agentic capabilities

Muse Spark 1.1 can orchestrate multi-agent systems to optimize end-to-end latency. As a main agent, it gathers context, creates plans, and delegates execution across parallel subagents. The model actively manages its 1 million token context window, remembering actions and retrieving information from earlier work.

For computer use, the model understands when to automate tasks versus direct interface interaction. It writes scripts for faster automation and generates batches of actions at each step rather than reasoning through every desktop click individually.

Coding performance improved substantially on real-world tasks involving large, complex codebases. The model can diagnose bugs, implement enterprise-grade features, and execute large code migrations. Meta developers and researchers now use Muse Spark 1.1 daily for internal development workflows.

Multimodal and safety features

The model excels in perception and multimodal reasoning, particularly when visual understanding and action execution must work together. It can extract photos from smartphone video, reason about products, and operate browsers to create Facebook Marketplace listings autonomously.

Meta conducted extensive safety evaluations following its Advanced AI Scaling Framework. Across chemical, biological, cybersecurity, and loss of control risk categories, Muse Spark 1.1 operates within safe margins with strong resistance to jailbreaks and indirect attacks.

Developer access and industry response

Developers can now access Muse Spark 1.1 through Meta's new Model API in public preview. The model is also available in "Thinking" mode in the Meta AI app and on meta.ai.

Early partners praised the model's comprehensive capabilities. "What's most impressive about Muse Spark is how much it packs into one model," said Amjad Masad, CEO of Replit, highlighting its million-token context, multimodal support, and coding abilities.

Meta indicated it has more capable models in training and plans to share future developments.