Meta AI

The Open-Source Gambit

Founded: 2013 (FAIR lab); Meta Platforms Inc. founded 2004 | HQ: Menlo Park, CA | Key People: Mark Zuckerberg (CEO), Alexandr Wang (Chief AI Officer), Joelle Pineau (VP AI Research) | Parent: Meta Platforms ($1.5T+ market cap) | Flagship Models: Muse Spark, Llama 4, Llama 4 Maverick, Llama 4 Scout

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. On April 9, 2026, Meta released Muse Spark—the first model from Meta Superintelligence Labs under Wang's leadership, representing a strategic shift away from the Llama brand for frontier releases. 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. Meta is now developing open-source versions of frontier models under the codenames "Avocado" and "Mango," signaling an intent to push the open-source frontier further and compete directly with proprietary models from OpenAI and Anthropic. No new Llama model was released in the April–May 2026 window. The company announced plans to fully automate its advertising business using AI by the end of 2026—a potentially transformative shift for Meta's $130+ billion annual ad revenue and a powerful proof-of-concept for enterprise AI automation at scale. The controversial Meta AI policy for WhatsApp went live in May 2026, enabling AI-assisted features across the messaging platform despite user privacy concerns.

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. The Avocado and Mango initiatives suggest Meta is not content to lag on frontier performance and is investing aggressively to close the gap with OpenAI and Anthropic in open-source form. The company's greatest asset remains distribution: Meta AI reaches over 500 million users through Facebook, Instagram, WhatsApp, and Messenger, giving it a consumer AI footprint second only to Google. The full ad automation initiative—if successful—would be the most consequential enterprise AI deployment to date, proving that AI can autonomously manage a $130+ billion revenue stream. However, the WhatsApp AI policy controversy highlights the tension between aggressive AI deployment and user trust. Meta's frontier model capabilities have not consistently matched the performance of proprietary alternatives, and the benchmark controversies from Llama 4 linger.

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. Watch for the Avocado and Mango model releases as potential game-changers for open-source AI capability. The full ad automation initiative is the most ambitious enterprise AI deployment announced to date—if your organization is in advertising, marketing, or media, this will reshape competitive dynamics by the end of 2026. However, Meta's frontier models have not consistently matched the performance of proprietary alternatives from OpenAI or Anthropic. The WhatsApp AI policy controversy should remind leaders that aggressive AI deployment carries reputational risk. 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.