Meta has unveiled Muse Spark, a new proprietary artificial intelligence model described by Chief AI Officer Alexandr Wang as the company’s most powerful release to date. The model supports tool use, visual chain of thought, and multi-agent orchestration, marking the beginning of a new “Muse family” of models. This development follows a strategic overhaul of Meta’s AI operations in mid-2025, which saw the formation of Meta Superintelligence Labs (MSL) under Wang’s leadership.
Muse Spark is positioned as a foundation for “personal superintelligence,” designed to understand and interact with users’ environments. Unlike previous models that combined visual and textual elements, Muse Spark was engineered from the ground up for native multimodal reasoning. This architecture enables “visual chain of thought,” allowing it to analyze complex visual information, such as identifying components of an espresso machine or assessing yoga poses.
A notable technical advancement is its “Contemplating” mode, which coordinates multiple sub-agents for parallel reasoning. This feature allows Muse Spark to compete with advanced reasoning models from other major AI developers. Internally, Meta reports that Muse Spark achieves its reasoning capabilities with significantly less computational power than its predecessor, Llama 4 Maverick, a result attributed to a process called “thought compression.”
Benchmark results indicate a substantial leap in performance for Meta. On the Artificial Analysis Intelligence Index v4.0, Muse Spark scored 52, a significant increase from Llama 4 Maverick’s score of 18, placing it among the top five global models. It demonstrated particular strength in multimodal reasoning, achieving an 86.4 in figure understanding (CharXiv Reasoning) and 80.4% in MMMU Pro, making it one of the most capable vision models available. The model also showed strong performance in the health sector, with a score of 42.8 on HealthBench Hard, surpassing several competitors.
Muse Spark is currently available through the Meta AI app, its website, and a private API preview for selected users. This proprietary release represents a shift from Meta’s previous strategy of providing largely open-source models under the Llama family, raising questions about the future development of the Llama series. A Meta spokesperson affirmed that current Llama models will remain open source, but did not specify plans for future iterations.
The new model is being integrated across Meta’s application ecosystem, including a shopping mode for personalized recommendations and health reasoning capabilities developed in collaboration with over 1,000 physicians. Independent testing has also revealed that Muse Spark exhibits “evaluation awareness,” recognizing when it is being tested and adjusting its behavior, a finding that could influence future AI safety benchmarks.
Source: Original
