Meta has introduced Muse Spark, the first in its family of multimodal reasoning models developed by Meta Superintelligence Labs. This new model supports tool-use, visual chain of thought, and multi-agent orchestration, marking the first product of Meta's ground-up AI overhaul. Muse Spark achieves competitive performance in multimodal perception, reasoning, health, and agentic tasks, with special emphasis on visual STEM questions and entity recognition.

The model's Contemplating mode orchestrates multiple reasoning agents in parallel, enabling it to compete with extreme reasoning modes of frontier models. This approach has achieved 58% on Humanity's Last Exam and 38% on FrontierScience Research benchmarks. Meta claims they can reach these capabilities with over an order of magnitude less compute than their previous model, Llama 4 Maverick, through improvements to model architecture, optimization, and data curation.

Available now at meta.ai and the Meta AI app with a private API preview, Muse Spark represents Meta's strategic investment across the entire AI stack, from research to infrastructure including the Hyperion data center. The company conducted extensive safety evaluations following the Advanced AI Scaling Framework, confirming strong refusal behavior in high-risk domains and safe margins across all measured risk categories.