Meta has officially launched Muse Spark, its first AI model from the newly formed Meta Superintelligence Labs, positioning it as a step toward "personal superintelligence." The model, announced on April 8, 2026, is now live on meta.ai and the Meta AI app, with a private API preview rolling out to select partners. A broader rollout to WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban AI glasses is planned for the coming weeks.
Muse Spark represents a strategic pivot for Meta, moving away from purely open-source models like Llama toward a closed, proprietary architecture. The company rebuilt its AI stack over nine months under the Superintelligence Labs banner, led by Chief AI Officer Alexandr Wang following Meta's $14 billion acquisition of Scale AI. Meta claims the new pretraining stack achieves capabilities similar to its previous Llama 4 Maverick model using "over an order of magnitude less compute."
A central feature is the new "Contemplating" mode, which orchestrates multiple AI agents to run in parallel for complex reasoning tasks. In this configuration, Muse Spark achieved scores of 58% on Humanity's Last Exam and 38% on the Frontier Science Research benchmark, performance Meta says allows it to compete with the extreme reasoning modes of rivals like Google's Gemini Deep Think and OpenAI's GPT Pro.
The model is natively multimodal, processing and generating text and images from the ground up, and includes capabilities for tool use and managing sub-agents. On specific benchmarks, Muse Spark scored 42.8% on the HealthBench Hard benchmark, outperforming GPT-5.4 (40.1%) and Gemini 3.1 Pro (20.6%). It also led on agentic search (DeepSearchQA) with a score of 74.8.
Meta is heavily emphasizing health applications, stating it collaborated with more than 1,000 physicians to curate training data to improve medical reasoning. The company argues health and wellness queries are a top reason people use AI and a key area for differentiation. However, analysts warn this focus could attract regulatory scrutiny in regions like the U.S. and EU concerning health advice and data privacy.
Despite strong showings in niche areas, overall benchmarks indicate Gemini 3.1 Pro still holds a lead in core categories like abstract reasoning (ARC AGI 2) and coding (LiveCodeBench Pro). Meta's blog acknowledges current performance gaps in long-horizon agentic systems and coding workflows.
Following the announcement, Meta's stock jumped approximately 9% during the trading day, closing up 6.5% at $612.42, as investors bet on a stronger AI strategy driving new revenue.
