Meta AI
The Open-Source Gambit
The Origin Story
Meta's AI journey began in 2013 when the company, then called Facebook, hired Yann LeCun—one of the three godfathers of deep learning—to found the Facebook AI Research lab (FAIR). LeCun's convivial presence and scientific credibility attracted top-tier research talent and established FAIR as one of the world's premier AI research centers, producing foundational work on convolutional neural networks, self-supervised learning, and computer vision. For years, FAIR operated as an academic-style lab within a social media company, publishing openly and contributing heavily to the open-source PyTorch framework, which became the industry standard for deep learning. The strategic pivot came in 2023 when Meta, facing competitive pressure from OpenAI and Google, decided to release its Llama model weights to the research community—a decision that inadvertently triggered the open-source AI revolution and fundamentally altered the competitive dynamics of the AI industry.
Key Milestones
The release of Llama 1 in February 2023 was initially a modest affair—a research-only license for a 65-billion-parameter model. When the weights leaked online via BitTorrent, the open-source community seized on them, creating fine-tuned variants that rivaled proprietary models. Meta took the lesson: Llama 2, released in July 2023, was explicitly open-source with a commercial license, and Llama 3 in April 2024 pushed performance further with models ranging from 8 billion to 70 billion parameters. Llama 4, released in April 2025, represented Meta's most ambitious effort yet, introducing a mixture-of-experts architecture with models scaling up to the reported Behemoth configuration of over one trillion parameters. However, the launch was marred by controversy: allegations emerged that Meta had optimized Llama 4's benchmark submissions in ways that did not reflect real-world performance, and Yann LeCun departed the company amid frustration over the direction of AI strategy, publicly criticizing the benchmark gaming. LeCun subsequently raised over $1 billion for a new AI startup. In mid-2025, Meta acquired Scale AI for approximately $15 billion and installed its founder, 28-year-old Alexandr Wang, as Chief AI Officer. Wang was tasked with creating Meta Superintelligence Labs, a new division focused on developing artificial general intelligence. The division reportedly debated whether to abandon the Behemoth model entirely in favor of fresh architectures. By April 2026, Meta released Muse Spark as a successor to Llama, signaling a strategic shift. Meta's capital expenditure budget for 2026 stands at $115 to $135 billion, nearly double the $72 billion spent in 2025, making it the largest single-company AI infrastructure investment in history.
Current Position
Meta's open-source strategy has made Llama the most widely deployed model family in the world, with hundreds of thousands of developers building on Llama-based architectures. The strategy serves Meta's core interest: commoditizing the model layer to ensure no competitor can extract monopoly rents from AI infrastructure that Meta depends upon. However, the company's frontier model capabilities have lagged behind OpenAI, Google, and Anthropic. The Llama 4 benchmark controversy and LeCun's departure damaged credibility. Wang's appointment signals a willingness to spend aggressively—Meta's AI budget now rivals the GDP of small nations—to close the gap. The company's greatest asset is distribution: Meta AI reaches over 500 million users through Facebook, Instagram, WhatsApp, and Messenger, giving it a consumer AI footprint second only to Google.
What Leaders Should Know
Meta's open-source models are the pragmatic choice for organizations that want AI capabilities without vendor lock-in. If your company needs to run models on-premises, customize weights for domain-specific tasks, or avoid per-token API costs at scale, Llama-based solutions offer the most mature ecosystem. However, Meta's frontier models have not consistently matched the performance of proprietary alternatives from OpenAI or Anthropic. The benchmark controversies should give procurement teams pause when evaluating published metrics. Meta's massive infrastructure investment signals long-term commitment, but the leadership turbulence—LeCun's departure, Wang's new mandate—introduces strategic uncertainty. For C-suite leaders, the lesson is clear: use Meta's open-source models for cost-effective deployment, but evaluate proprietary models for mission-critical applications requiring frontier performance.