Meta Launches Muse Spark as AI Adoption Surges Despite Performance Gaps

Meta Platforms has released Muse Spark, the first artificial intelligence model from a specialized team assembled to close the gap with competitors, as pressure mounts on tech giants to justify massive AI investments.

The model offers practical consumer features including calorie estimation from meal photographs and visual placement tools for home decoration. A "Contemplating Mode" runs multiple agents simultaneously to enhance reasoning capabilities, positioning Muse Spark against extended thinking systems from Google and OpenAI. Meta suggested the mode could coordinate vacation planning by having one agent draft itineraries while another researches family activities.

The launch arrives amid broader evidence of accelerating AI adoption across the economy. Stanford University's 2026 AI Index, released this week, found that 88 percent of organizations used AI for at least one business function last year, up ten percentage points from 2024. Among AI users, 79 percent reported regular use of generative AI in business operations, compared to 71 percent the previous year.

Yet the same research reveals persistent performance limitations. Human scientists continue to outperform the best AI agents on complex tasks, according to the Stanford report. A separate study of Google's AI Overviews found that while accuracy improved from 85 percent under Gemini 2 to 91 percent under Gemini 3, more than half of accurate responses lacked proper grounding—linking to websites that did not fully support the information provided.

"Whether or not this explosive growth is meaningful is hotly debated," said computer scientist Arvind Narayanan at Princeton University. "My view is that it is happening too fast, without giving scientific norms time to adjust, and so the quality of research has taken a nosedive."

Financial institutions are among the most aggressive adopters. Citigroup reported using AI to reduce account opening document review from over an hour to 15 minutes, while also accelerating legacy system migrations through automated coding and testing. Tim Ryan, the bank's head of technology, described the tools as central to productivity improvements following years of infrastructure investment.

(The Stanford index tracks and aggregates data from multiple surveys and studies to provide an annual snapshot of AI development and deployment. The amount of AI-generated content published online surpassed human-authored articles in November 2024, according to the report.)

Meta's entry into advanced reasoning models reflects intensifying competition in AI capabilities. The performance gap between leading U.S. and Chinese models has effectively closed, with Anthropic's top system holding just a 2.7 percentage point advantage over the best Chinese competitor as of March 2026. That margin has fluctuated repeatedly since DeepSeek's R1 briefly matched American models in early 2025. Some U.S. companies have alleged that Chinese labs achieved progress through adversarial distillation—training models on competitor outputs—though evidence supporting the extent of this practice has not been publicly released.