How Meta Won Open Source AI, Then Threw It Away

Meta's rise and fall as the champion of open-source AI is traced through the Llama model series. Starting with the surprise 2023 Llama release that democratized access to frontier AI, Meta built enormous developer trust through Llama 2 and 3. But the Llama 4 launch in April 2025 unraveled that trust: Meta submitted a specially tuned, unreleased model to public leaderboards, skipped the technical paper that accompanied every prior release, and published benchmark claims that independent testing couldn't replicate. The fallout included the sidelining of the entire GenAI team, the departure of Turing Award winner Yann LeCun, a $14.3B investment in Scale AI, and the hiring of Alexander Wang as Chief AI Officer. LeCun later confirmed benchmarks were manipulated and attributed the failure to prioritizing safe, deadline-driven choices over genuine research. Meanwhile, DeepSeek, Alibaba, and Mistral have emerged as credible alternatives, and Meta's flagship Behemoth model still hasn't shipped.