Meta AI has quietly integrated with Google Gemini, and at least five variants of Avocado are still in internal testing.

According to 1M AI News monitoring, Meta is internally running parallel tests of at least 5 Avocado model variants:

1. Avocado 9B: a small version with 9 billion parameters
2. Avocado Mango: a multimodal variant marked as “agent” and “sub-agent,” supporting image generation
3. Avocado TOMM: short for “Tool of Many Models,” based on Avocado’s multi-model orchestration architecture
4. Avocado Thinking 5.6: the latest reasoning model
5. Paricado: a pure text dialogue model

Multiple product lines running in parallel means Meta hasn’t yet decided the final shipping form. Even more worth paying attention to is that some users’ requests for Meta AI have already been processed in practice by Google’s Gemini models, and are no longer stuck in the “internal discussions” stage reported earlier. Avocado had previously fallen behind the frontier models of Google, OpenAI, and Anthropic due to internal testing, and its release timeline has been pushed back from March to at the earliest May.

In addition, Avocado is expected to be released as a closed-source product—Meta’s clear pivot away from the open-source path of the Llama era. Under Zuckerberg’s goal of pursuing superintelligence, Meta’s AI strategy is shifting from an “open-source benchmark” to “performance first.”