Meta has officially entered the frontier “superintelligence” race with the launch of Muse Spark, its most powerful multimodal AI model to date. Announced on April 8 by Mark Zuckerberg, Muse Spark is the first product from the newly formed Meta Superintelligence Labs (MSL), led by former Scale AI CEO Alexandr Wang.
The model represents a “ground-up” architectural shift for Meta, moving away from the open-weights Llama series toward a proprietary, natively multimodal system designed for “personal superintelligence.”
1. Key Innovation: “Visual Chain-of-Thought”
Unlike previous models that “stitched” a vision layer onto a text model, Muse Spark was built to integrate visual information into its internal logic.
- Multimodal Reasoning: It doesn’t just recognize objects; it understands dynamic environments. It can identify components of complex machinery or correct a user’s yoga form via side-by-side video analysis.
- Perception at Scale: You can snap a photo of a shelf and ask Meta AI to “rank these snacks by protein content,” or scan a diagram to have the AI explain the physics involved.
- Visual Grounding: The model can annotate images and videos in real-time, making it highly effective for tutorials and “how-to” tasks.
2. The “Contemplating” Mode
To compete with the “Deep Think” and “Pro” modes of its rivals, Meta introduced Contemplating Mode, an agentic reasoning framework.
- Parallel Agents: Instead of a single long chain of thought, Muse Spark orchestrates multiple “sub-agents” that reason in parallel, collaborate, and synthesize a final answer.
- Scientific Dominance: This mode allows Muse Spark to lead in extremely hard scientific benchmarks, scoring 58% on Humanity’s Last Exam (HLE), outperforming both GPT-5.4 Pro and Gemini 3.1 Pro.
- Efficiency: Through a process called “thought compression,” Meta claims Muse Spark achieves this reasoning using 10x less compute than its previous Llama 4 Maverick flagship.
3. Specialized Modes: Shopping and Health
Meta is leveraging its unique social data and creator ecosystem to power two highly specialized experiences:
- Shopping Mode: The AI identifies brands, styling choices, and products across Instagram and Threads, providing personalized recommendations and effectively turning every post into a shoppable interaction.
- Health Reasoning: Developed in collaboration with over 1,000 physicians, Muse Spark can analyze nutritional content from food photos and provide structured health insights based on charts or medical visuals.
4. Benchmark Performance: A New Leader in Health
While Muse Spark leads in health and vision, it still trails in coding and abstract problem-solving.
| Benchmark | Muse Spark | GPT-5.4 Xhigh | Gemini 3.1 Pro |
| HealthBench Hard | 42.8 (Winner) | 40.1 | 20.6 |
| GPQA (PhD Reasoning) | 89.5 | 92.8 | 94.3 |
| Terminal-Bench (Coding) | 59.0 | 75.1 | 68.5 |
| ARC-AGI-2 (Abstract) | 42.5 | 76.1 | 76.5 |
5. Availability and Privacy
- Where to try it: Available now at meta.ai and via the Meta AI app. It is currently rolling out to WhatsApp, Instagram, and Facebook in the US, with AI glasses integration coming shortly.
- Interactive UI: You can prompt Muse Spark to build “visual artifacts,” such as generating a playable Sudoku game from a photo or launching a whimsical flight simulator.
- API Access: A private preview is available via API for select partners, though Meta has indicated it hopes to open-source future (smaller) versions of the Muse series.
“Muse Spark is the first step on our scaling ladder,” Meta noted. “It’s the difference between an AI that waits for you to explain the world and one that can simply look at the world with you.”