Meta Superintelligence Labs introduced Muse Spark, the first in a new family of natively multimodal reasoning models with tool-use, visual chain of thought, and multi-agent orchestration.

Muse Spark achieves competitive results in multimodal perception, health reasoning curated with over 1,000 physicians, and agentic tasks; its Contemplating mode reaches 58% on Humanity’s Last Exam and 38% on FrontierScience Research, matching extreme reasoning modes in Gemini Deep Think and GPT Pro (https://ai.meta.com/blog/introducing-muse-spark-msl/; https://openai.com/index/introducing-o1/). The pretraining stack delivers equivalent capabilities with over 10x less compute than Llama 4 Maverick per fitted scaling laws.

Meta's blog post details three scaling axes—pretraining, reinforcement learning, and test-time reasoning—yet omits explicit parameter counts, total FLOPs, or head-to-head benchmarks versus o1 and Claude 3.5 Sonnet; coverage also underplays how visual CoT directly extends o1-style internal chains into grounded multimodal settings (https://openai.com/index/introducing-o1/; https://www.anthropic.com/news/claude-3-5-sonnet).

Applications center on interactive health explanations, nutritional analysis, muscle activation maps, and annotated appliance troubleshooting, synthesizing Meta's Llama efficiency gains with physician-sourced data to target personal superintelligence use cases left unaddressed in prior text-only releases.