Meta is pushing Muse Spark toward agentic with a 1.1 release.

Meta Superintelligence Labs releases Muse Spark 1.1, a multimodal model for agentic tasks, tool use, and multi-agent systems via the new Meta Model API.

Meta Superintelligence Labs is releasing Muse Spark 1.1, a revision of its multimodal reasoning model geared toward agentic tasks. The announced advancements focus on tool and computer use, code, and multimodal understanding.

The model is designed to orchestrate multi-agent systems. As a primary agent, it gathers context, establishes a plan, and distributes execution among parallel sub-agents; as a sub-agent, it sticks to its task and knows when to report back to the primary. Meta claims it generalizes zero-shot to new native tools, MCP servers, and skills, and actively manages a one-million-token context window, retrieving information from distant steps and compacting it without losing useful milestones.

Regarding computer use, the publisher describes a model that arbitrates between writing a script when automation is faster and clicking when direct interaction is simpler, all while generating batches of actions at each step. It maintains continuity over long sessions and navigates unfamiliar interfaces with minimal human intervention. On the code front, Meta highlights bug diagnosis and correction, adding features to complex systems, and large migrations, with good adaptation to common agentic coding harnesses (planning mode, delegation to sub-agents, context compaction). The multimodal aspect targets cases where perception and action go hand-in-hand, such as listing an item on Facebook Marketplace from a smartphone video.

Meta is accompanying the release with a public preview of its new Meta Model API, making the model accessible to developers. It is also offered in "Thinking" mode within the Meta AI app and on meta.ai. On security, the publisher states it conducted its evaluations according to its Advanced AI Scaling Framework and places the model within safe margins for chemical and biological risks, cybersecurity, and loss of control.