According to 1M AI News monitoring, Meta is internally parallel testing at least 5 Avocado model variants:
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," based on Avocado's multi-model scheduling architecture
4. Avocado Thinking 5.6: the latest inference model
5. Paricado: a text-only dialogue model
Running multiple product lines in parallel means Meta has not yet finalized the shipping form. Of more significance, some user requests to Meta AI have actually been processed through Google's Gemini model, moving beyond the previously reported "internal discussion" stage. Avocado, previously lagging in internal testing compared to cutting-edge models from Google, OpenAI, and Anthropic, has seen its release date pushed back from March to as early as May.
Furthermore, Avocado is expected to follow a closed-source release, marking Meta's clear shift from the Llama era's open-source roadmap. As Mark Zuckerberg pursues the goal of superintelligence, Meta's AI strategy is transitioning from being an "open-source benchmark" to "performance-first."
