Meta is making a dramatic pivot in its artificial intelligence strategy, abandoning the open-source approach that built its developer community and betting billions on proprietary models instead. The company's new Muse Spark model, developed by the newly formed Meta Superintelligence Labs (MSL), ranks fourth on the Artificial Analysis Intelligence Index with a score of 52, trailing only Gemini 3.1 Pro, GPT-5.4, and Claude Opus 4.6 . This represents a stunning recovery from the company's Llama 4 debut just a year ago, which scored only 18 points and damaged Meta's credibility through benchmark manipulation .

What Happened to Meta's Open-Source Leadership?

For nearly three years, Meta built a loyal global developer community by releasing its Llama family of models as open-source software, allowing anyone to download, modify, and deploy them freely. That era has ended. Muse Spark is Meta's first frontier-class model release since April 2025, and it is not available as open weights . Instead, the model runs exclusively through the Meta AI app, on meta.ai, and via a private application programming interface (API) preview for select partners . This represents a fundamental break from the strategy that made Meta unique in the competitive AI landscape.

The shift reflects a calculated business decision. Mark Zuckerberg's vision of giving every person their own personal AI agent requires maintaining tight control over the underlying technology . When you plan to embed AI agents into billions of devices and apps, keeping the model proprietary ensures you can monetize the service and prevent competitors from using your research directly. Meta stated that existing Llama models would remain available as open source, but left the future of the Llama family deliberately ambiguous .

How Is Meta Rebuilding Its AI Division?

The transformation began in summer 2025 when Zuckerberg restructured the entire AI division following Llama 4's failure. He brought in Alexandr Wang, the 29-year-old co-founder of Scale AI, as Chief AI Officer, backed by a $14.3 billion investment in Scale AI for a 49 percent non-voting stake . Wang leads Meta Superintelligence Labs, which rebuilt the entire AI stack from scratch, including architecture, data pipelines, and training infrastructure . The results speak for themselves: Meta achieved comparable capabilities to its previous flagship model using over an order of magnitude less computing power .

To attract top talent from OpenAI, Google, and Apple, Zuckerberg reportedly offered compensation packages worth up to $200 million over four years . Meanwhile, Yann LeCun, Meta's previous AI chief, departed to build his own startup, AMI Labs in Paris, with billions in investment .

Steps to Understanding Meta's Infrastructure Buildout

• Immediate Investment: Meta announced capital expenditures of $115 to $135 billion for 2026, nearly double the 2025 figure, with plans to invest at least $600 billion in US data centers and AI infrastructure by 2028 .
• Prometheus Data Center: A 1-gigawatt facility scheduled to come online in 2026, representing a single massive computing cluster dedicated to training and running AI models .
• Hyperion Mega-Cluster: Located in Richland Parish, Louisiana, this 2,250-acre site will cost approximately $10 billion and deliver 5 gigawatts of compute capacity with over one million graphics processing units (GPUs), roughly equivalent to a quarter of Manhattan's land area .

This infrastructure spending dwarfs what most companies can afford. To put it in perspective, Meta is essentially building multiple power plants worth of computing capacity dedicated entirely to AI development and deployment .

What Makes Muse Spark Different From Previous Meta Models?

Muse Spark is natively multimodal, meaning it can process text, images, and other data types simultaneously . The model supports tool use, allowing it to interact with external software and services, and includes visual chain of thought reasoning, which helps it explain its thinking process through visual representations . Most notably, a new "Contemplating Mode" enables multi-agent orchestration, where several AI sub-agents work in parallel on a single question, helping the model compete with advanced reasoning modes from competitors like Gemini Deep Think and GPT Pro .

Meta invested heavily in health reasoning capabilities, working with over 1,000 physicians to curate training data for medical knowledge . On the Arena.ai leaderboard, formerly known as LMArena, Muse Spark currently trails only Claude 4.6 in text and vision tasks, though it lags significantly in coding workflows . For Meta's target audience of consumers on Facebook, Instagram, WhatsApp, and the Meta AI app, this performance profile is acceptable .

The model will roll out gradually across Meta's ecosystem. Muse Spark now powers the standalone Meta AI app and desktop website, with deployment coming to Facebook, Instagram, WhatsApp, and Messenger in the coming weeks, as well as Meta's Ray-Ban AI glasses . A new Shopping mode will help users find clothing and home decor by drawing from styling inspiration and brand storytelling already happening across Meta's apps .

Why Did Meta Acquire Manus, and What Does It Mean for Business Users?

In late December 2025, Meta closed an acquisition of Manus, a Singapore-based provider of autonomous AI agents, for over $2 billion . Manus originally emerged from a Chinese startup, but all remaining Chinese ties were severed as part of the deal . Unlike conventional chatbots that respond to individual prompts, Manus agents handle multi-step tasks largely autonomously, from market research and coding to data analysis .

The acquisition gives Meta an immediately profitable business. Manus already sells AI agents on a subscription basis to small and medium-sized businesses and recently reached an annualized revenue run rate exceeding $125 million . This means Meta essentially acquired a functioning business generating real revenue overnight, not just technology .

Strategically, Meta is targeting the small and medium-sized business segment, which already represents one of Meta's most important advertising customer bases on Facebook, Instagram, and especially WhatsApp Business . Manus's agentic capabilities map seamlessly onto the Meta Business Suite, where small businesses manage content calendars, inboxes, ads, and analytics . An execution agent could automate all of that end-to-end, transforming WhatsApp Business, which has over two billion users, into something resembling an autonomous digital employee for millions of small companies . This mirrors what WeChat has long offered in China .

What Is Zuckerberg's Broader Vision for Personal AI Agents?

Zuckerberg's ultimate goal is what he calls "Personal Superintelligence": every single person should have their own AI agent that thinks, plans, communicates, and acts on their behalf in everyday life . According to a Wall Street Journal report from March 2, 2026, Zuckerberg is already using a personal AI agent himself and plans to roll out this capability to everyone inside and outside the company . Muse Spark is the first building block toward this vision, with deployment planned through Meta's apps and, critically, through Meta's augmented reality glasses developed in partnership with Ray-Ban .

The dual strategy is clear: Muse Spark delivers the personal agent for consumers, while Manus provides the business agent for the long tail of small companies . Together, they position Meta to capture value across both consumer and business segments of the emerging AI agent economy.

After a year of public setbacks and internal turmoil, Meta has re-entered the frontier AI race with a completely new philosophy and unprecedented financial commitment. The price of this comeback is the open-source identity that made Meta unique, replaced by a proprietary, vertically integrated approach designed to keep control over the personal AI agents of the future .

" }