Project Avocado & Mango: Inside Meta’s Six-Month Sprint to Rebuild Its Generative AI Lead

Meta Platforms’ newly formed Superintelligence Labs has produced its first in-house AI models, Chief Technology Officer Andrew Bosworth said at a World Economic Forum briefing in Davos, calling the early results “very good.” The update marks a visible milestone in Meta’s renewed push to regain momentum in generative AI following sweeping leadership changes and heavy investment over the past year.

Bosworth said the models were developed in roughly six months but remain too immature for public release. He stressed that “there’s a tremendous amount of work to do post-training” before they can be packaged into products for internal teams or consumers. He did not specify which model families had been delivered, though earlier media reports have pointed to projects codenamed Avocado for text and Mango for image and video.

Product and Strategy

The progress comes amid a major reorganisation at Meta that elevated AI to a central strategic priority, including the creation of the Superintelligence Labs unit, aggressive hiring and significant infrastructure spending through 2025. Bosworth framed 2026–27 as a crucial period for translating recent research advances into consumer-facing products, adding that the company is already seeing encouraging returns from last year’s investments.

Meta is also aligning its consumer hardware and services more closely with its AI roadmap. The company recently paused the international rollout of its AI-enabled Ray-Ban Display glasses to prioritise fulfilling U.S. demand, a move Bosworth described as reflecting supply constraints and go-to-market focus while the underlying technology continues to stabilise.

The effort carries reputational and competitive stakes. Meta has faced criticism over earlier models such as Llama 4, while rivals including Alphabet and OpenAI have continued to move quickly. Bosworth acknowledged the gap but said early results from the new lab suggest Meta can close it with further engineering and productisation.

Meta Superintelligence Bet

The Superintelligence Labs push has been backed by an aggressive talent drive, with reports of senior researchers hired from rival AI labs on lucrative packages, alongside major spending on compute and data-centre capacity.

If Meta can successfully turn these early models into reliable, scalable features, it could accelerate AI integration across its social, messaging and VR/AR products. Still, Bosworth’s emphasis on extensive post-training work underscores that cutting-edge models must clear hurdles around safety, latency, alignment and product engineering before they can meaningfully shift the competitive ecosystem.