Quick Facts
- Meta launched Muse Spark, its first proprietary AI model from newly formed Superintelligence Labs, scoring 52 on AI benchmarks and ranking in top 5 globally
- The model excels in health applications with a 42.8 score on HealthBench Hard, outperforming GPT-5.4 and Gemini 3.1 Pro
- Meta stock surged 9% Wednesday as analysts see potential for significant monetization through AI-powered shopping experiences
Meta launched Muse Spark Wednesday, marking the company’s first proprietary artificial intelligence model since forming Meta Superintelligence Labs. The release represents a strategic shift away from the open-source approach that defined the company’s Llama series.
Muse Spark scores 52 on the Artificial Analysis Intelligence Index v4.0, placing it fifth globally behind GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.6. The model demonstrates particular strength in multimodal understanding, scoring 80.5% on MMMU-Pro benchmarks.
The model dominates health-related applications with a 42.8 score on HealthBench Hard, substantially ahead of competitors including GPT-5.4 at 40.1 and Gemini 3.1 Pro at 20.6. Meta collaborated with over 1,000 physicians to develop specialized training data for medical queries.
Strategic Reorganization
Chief AI Officer Alexandr Wang, who joined Meta through the company’s $14.3 billion investment in Scale AI, led the nine-month rebuild of Meta’s AI infrastructure. Wang stated on X that his team has more models planned, with some returning to open-source releases.
The launch follows Meta’s formation of Superintelligence Labs after Llama 4’s disappointing reception among developers. CEO Mark Zuckerberg overhauled the company’s AI operations in summer 2025, bringing Wang aboard as the first chief AI officer.
Market Response
Meta shares jumped 9% Wednesday, marking the sharpest gain since January. Mizuho maintained an Outperform rating with an $850 price target, citing the model’s potential to drive monetization through AI-powered shopping experiences.
The firm highlighted how usage data from shopping applications could enhance ad targeting in Meta’s core newsfeed products. Meta expects AI-related capital expenditures between $115 billion and $135 billion in 2026, nearly double last year’s spending.
Competitive Position
The proprietary model marks Meta’s departure from open-source leadership as Chinese competitors gained ground. Chinese models from Alibaba and DeepSeek accounted for 41% of downloads on Hugging Face by late 2025, while U.S. models represented 35% of global Llama deployments.
Muse Spark introduces “contemplating mode,” orchestrating multiple reasoning agents in parallel to compete with Gemini Deep Think and GPT Pro for complex scientific tasks. Zuckerberg described the model as the “first milestone” toward personal superintelligence on Threads.
This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.