Muse Spark is the inaugural multimodal reasoning model in the Muse series developed by Meta Superintelligence Labs (MSL). Engineered as the foundation for personal superintelligence, Muse Spark features thought-compressed reinforcement learning, natively integrated visual perception, multi-agent orchestration, and specialized health and scientific reasoning—scaling across Meta AI (meta.ai), Ray-Ban Meta smart glasses, WhatsApp, Instagram, Facebook, and Threads.

Marking a major architectural transition following the Llama series, Meta Superintelligence Labs rebuilt its pre-training and reinforcement learning stack from the ground up. Muse Spark focuses on maximizing compute efficiency through test-time computation and thought compression, achieving the same capability milestones as prior architectures (such as Llama 4 Maverick) with over an order of magnitude less compute.

Architecture & Technical Profile at a Glance

Feature / Dimension	Muse Spark Technical Profile	Core Capability Advantage
Developer / Lab	Meta Superintelligence Labs (MSL)	Built as the initial foundation for the multi-tier Muse model family.
Model Type	Natively Multimodal Reasoning Model	Joint visual chain-of-thought, text processing, and tool invocation.
Reasoning Engine	Contemplating / Thinking Mode	Step-by-step reasoning achieving 58% on Humanity’s Last Exam.
Compute Efficiency	Thought-Compressed Reinforcement Learning	Penalizes overthinking to compress multi-step logic into fewer tokens.
Agent Architecture	Parallel Multi-Agent Sub-Delegation	Launches concurrent sub-agents for multi-stage planning and research.
Ecosystem Deployment	Meta AI, Ray-Ban Meta Glasses, WhatsApp, IG, FB	Integrated across hardware and apps for real-time camera & voice queries.

Core Breakthroughs & Capabilities

• Native Multimodal Perception & Environmental Grounding: Built to perceive and reason over the physical world in real time rather than relying solely on text prompts. Users can scan physical products, analyze nutrition charts, or troubleshoot appliances with interactive visual annotations.

• Reinforcement Learning with Thought Compression: Utilizes an advanced RL optimization stack that applies length penalties during extended test-time reasoning. This causes the model to naturally compress its internal reasoning traces, allowing it to solve complex STEM and logic puzzles with significantly lower token consumption and inference latency.

• Parallel Multi-Agent Swarm Delegation: Meta AI can decompose high-level user requests (e.g., end-to-end travel planning or multi-source marketplace comparison) into parallelized tasks handled simultaneously by multiple sub-agents.

• Physician-Curated Health & Wellness Reasoning: Trained in collaboration with over 1,000 physicians, Muse Spark provides grounded, comprehensive health context, visual symptom checks, and interactive displays (such as muscle activation during exercises or dietary macro breakdowns).

• Real-Time Natural Voice & Multimodal Glasses Integration: Powers conversational interaction on the Meta AI app and Ray-Ban Meta smart glasses—allowing users to interrupt, switch languages fluidly, and query what their camera sees on the go.