Meta to Open Source Next AI Models, Shift Focus to AI

Meta plans to open source its next AI models, a move to accelerate development and growth in the AI space. This decision aligns with the company's shift in focus from the metaverse to AI, following recent setbacks and employee layoffs.

META is moving to open-source its next generation of AI models, including planned open-weight releases of two proprietary frontier models codenamed Avocado — a large language model — and Mango, a multimodal generator. The announcement follows Meta's April 5 release of Llama 4 Scout and Llama 4 Maverick, both using a mixture-of-experts architecture with native multimodal capabilities (text, image, video), which Meta claims outperform GPT-4o and Gemini 2.0 Flash on key benchmarks.

The Avocado and Mango open-source versions will ship with deliberate capability gaps versus their proprietary counterparts — including reduced parameter counts and omitted post-training steps — with safety concerns cited around cybersecurity code generation. The strategy reflects Meta's $115-$135 billion 2026 AI infrastructure budget, among the largest in the industry. Unlike cloud rivals, Meta has no direct cloud business to monetize compute; its bet is that AI-enhanced advertising across WhatsApp, Facebook, and Instagram will justify the spend via ad revenue lift.

The pivot back toward open-source reverses a December 2025 Bloomberg report that suggested META was moving to a closed-source model. The company's LlamaCon event on April 29 is expected to provide further details on the roadmap. Analysts see the strategy as a calculated hedge: releasing capable-but-not-frontier models drives developer ecosystem adoption and shapes industry standards, while the most powerful capabilities remain proprietary for Meta's consumer platforms.