Meta introduced a new AI model, Muse Spark, and upgraded its Meta AI assistant with multimodal and agent-based capabilities, marking its latest push toward more personalized AI systems.
The model, developed by Meta’s Superintelligence Labs, is the first in a new “Muse” series designed to incrementally scale performance. The company said Muse Spark is “small and fast by design” but capable of handling complex reasoning tasks across science, math and health.
Muse Spark now powers Meta AI across its app and web interface, introducing new “Instant” and “Thinking” modes and the ability to deploy multiple AI subagents simultaneously. For example, Meta said the assistant can break down a task like travel planning into parallel workflows, generating itineraries, comparing destinations and recommending activities at the same time.
The upgrade also emphasizes multimodal capabilities. Users can upload images or use their camera to query the system, such as identifying products or analyzing food items. Meta said it worked with physicians to improve responses to common health-related questions, reflecting growing demand for AI in health care guidance.
The release comes amid intensifying competition among major AI developers, including OpenAI, Google and Anthropic, all of which are racing to deploy more capable multimodal systems and agent-based tools. Industry efforts are increasingly focused on “agentic AI,” where systems can autonomously execute tasks rather than simply generate responses.
Meta is positioning its approach around social context and personalization. New features integrate content from its platforms, including Instagram and Facebook, to inform recommendations for shopping, travel and trending topics. The company said future versions will incorporate richer media such as videos and posts directly into AI responses.
The rollout begins in the U.S., with broader global availability planned in the coming weeks. Meta also said it will offer API access to select partners and may open-source future versions of the Muse models.
The announcement underscores Meta’s longer-term ambition to build what it calls “personal superintelligence” — AI systems that not only answer questions but understand users’ preferences and environments.