Meta plans Mango multimodal AI and Avocado LLM for 2026 pivot from Llama

If Mango and Avocado land roughly as described, Meta is formally exiting the “open-source hero” posture and joining OpenAI, Google and Anthropic in the closed frontier-model camp. A multimodal Mango aimed at high-quality image and video generation, paired with Avocado as a frontier LLM tuned for coding and world modeling, would give Meta a vertically integrated stack for everything from Instagram filters to agentic assistants. It also acknowledges that open Llama weights have armed competitors—including Chinese labs—with strong baselines while doing little to monetize Meta’s own infrastructure spend.

Strategically, the shift matters because it tightens the oligopoly at the very top of the capability curve. Instead of being the open alternative, Meta is now competing directly in the same premium API market as its largest rivals, likely with aggressive cross-subsidies from advertising. That raises the bar for smaller labs trying to stay close to the frontier without sovereign backing. At the same time, reports that Avocado is targeting GPT‑5 and Gemini 3 Ultra–class performance suggest Meta accepts that the next leap will require world-model-like capabilities and much deeper integration of perception, action and reasoning.