Meta AI Quietly Integrates Google Gemini, At Least 5 Avocado Variants in Internal Testing

AirdropBlackHole · 2026-03-30T12:45:16+00:00

Meta is testing five variants of the Avocado model, including multimodal and scheduling architectures, reflecting its shift to a closed-source strategy. The release has been delayed due to internal testing and competition with major AI models.

AirdropBlackHole

2026-03-30 12:45:16

Abstract generation in progress

According to monitoring by 1M AI News, Meta is conducting parallel internal tests of at least five variants of the Avocado model: 1. Avocado 9B: a smaller version with 9 billion parameters; 2. Avocado Mango: labeled as ‘agent’ and ‘sub-agent’, a multimodal variant supporting image generation; 3. Avocado TOMM: short for ‘Tool of Many Models’, a multi-model scheduling architecture based on Avocado; 4. Avocado Thinking 5.6: the latest version of the inference model; 5. Paricado: a pure text dialogue model. The parallel development of multiple product lines indicates that Meta has not yet determined the final shipping form. Notably, some user requests for Meta AI have already been processed using the Google Gemini model, moving beyond the previously reported ‘internal discussion’ stage. Avocado’s release has been delayed from March to as early as May due to its internal testing lagging behind the cutting-edge models of Google, OpenAI, and Anthropic. Additionally, Avocado is expected to adopt a closed-source release, marking a clear shift in Meta’s approach from the open-source path of the Llama era. Under Zuckerberg’s pursuit of superintelligence, Meta’s AI strategy is shifting from ‘open-source benchmark’ to ‘performance-first’.

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.