Meta Superintelligence Lab delivers first internal frontier AI models
TL;DR
Meta CTO Andrew Bosworth said on January 21, 2026 that the company’s new Superintelligence Labs has delivered its first high-profile AI models for internal use. The lab, formed less than six months ago, is developing text and multimodal models codenamed “Avocado” and “Mango.”
About this summary
This article aggregates reporting from 5 news sources. The TL;DR is AI-generated from original reporting. Race to AGI's analysis provides editorial context on implications for AGI development.
Race to AGI Analysis
Meta’s Superintelligence Labs crossing the “first internal models” milestone matters because it signals that the company’s high-stakes reorg is actually producing new systems, not just slideware. After a rocky Llama 4 cycle, Meta effectively rebooted its frontier program under Alexandr Wang with an explicit brief to push beyond open-source baseline models. Delivering Avocado and Mango internally in under six months shows that the new pipeline—talent, data, infra—is starting to hum.
This puts Meta back into the front pack of labs racing to build general-purpose systems with strong multimodal and agentic capabilities. Unlike OpenAI or Anthropic, Meta’s distribution edge is its installed base: billions of users across Facebook, Instagram, WhatsApp and its VR/AR platforms. Internal deployment lets it harden models on real product workloads—recommendations, assistants, creative tools—before deciding what, if anything, to release as open weights.
For the broader race to AGI, Superintelligence Labs is a reminder that the contest isn’t just about who has the single best model today, but who can repeatedly spin up new model generations and wire them into massive social products. If Meta can turn these internal models into compelling, sticky AI experiences at scale, it meaningfully shifts the competitive balance with OpenAI, Google and xAI.
