Meta AI Updates: April 7, 2026

1. Meta Shifts to Hybrid Open/Closed Strategy for Next-Generation AI Models

Meta. Axios reported that Meta is preparing to release the first AI models developed under Alexandr Wang (formerly Scale AI CEO, now leading Meta’s superintelligence team) and plans to eventually offer versions under an open-source license. In a strategic shift from Meta’s previous open-source-first approach, some of the largest new models — including codenames “Avocado” for text LLM and “Mango” for image/video — will initially remain proprietary to manage safety risks. Meta is targeting consumers rather than the enterprise market favored by OpenAI and Anthropic. Source

2. Meta AI Releases EUPE: Compact Vision Encoders Under 100M Parameters

Meta. Meta AI released EUPE (Efficient Universal Perception Encoder), a family of compact vision encoders under 100M parameters designed for on-device and edge deployment. EUPE distills knowledge from a 1.9-billion-parameter “proxy teacher” model — itself created by aggregating multiple domain-expert foundation vision encoders — down to efficient student models that match or exceed the performance of individual specialist encoders of similar size. Code and model weights are released under the FAIR Research License. Source