Meta unveils Muse Spark, its first new AI model since hiring Alexandr Wang

Key Points:

• Meta has launched Muse Spark, its first reasoning and multimodal AI model capable of step-by-step problem solving, using text and images, and coordinating multiple subagents to tackle complex tasks.
• Muse Spark is designed to be small and fast while handling intricate questions in science, math, and health, serving as a testbed before scaling to larger models in the same family.
• Benchmark tests show Muse Spark performs competitively but not dominantly, scoring slightly below top models like Gemini 3.1 Pro and GPT-5.4 on PhD-level reasoning, but outperforming rivals on a challenging health benchmark.
• The model's launch reflects Meta's major AI reorganization, including a $14.3 billion investment in Scale AI and the creation of the Meta Superintelligence Labs led by Alexandr Wang, alongside a new applied AI engineering unit to balance long-term research with product development.
• Meta reports significant improvements in AI architecture, optimization, and data curation, enabling Muse Spark to achieve comparable capabilities to previous models with substantially less computing power, marking the first step in a deliberate scaling strategy.