Peter Smith
Meta Platforms is actively testing a new generation of AI systems under the internal codename "Avocado." A leaked internal model selector, first reported by Wes Roth, reveals multiple variants currently under evaluation - signaling continued development across Meta's AI stack.
The Avocado lineup includes three distinct models under simultaneous testing:
- Avocado 9B - a 9-billion parameter efficiency-focused model
- Avocado Thinking 5.6 - a reasoning-oriented variant
- Paricado - a text-only conversational model
Meta appears to be exploring parallel optimization paths across efficiency, reasoning, and conversational performance rather than focusing on a single architecture.
The internal interface suggests the company is evaluating multiple configurations at once. This modular approach stands in contrast to the single-flagship model strategy that defined earlier generations of large language models. Instead of betting on one generalized system, Meta is building specialized AI tailored to different task types.
The emergence of the Avocado family underscores how rapidly evolving AI strategies are becoming central to competitive positioning in the tech sector.
This push into specialized architecture isn't happening in isolation. Recent work across Meta SAM 3.1 performance improvements demonstrates a clear focus on compute efficiency, while Meta's HyperAgents recursive systems point toward self-improving multi-agent frameworks as a longer-term research direction.
The presence of multiple model types within the same testing environment highlights a modular approach that aligns with broader industry trends.
At the same time, reasoning model evaluation challenges remain an open problem across the field - with Meta's own research showing that even high-scoring LLM judges are still vulnerable to reward hacking. Taken together, these developments reflect a company investing deeply in AI infrastructure while actively working through the limitations that come with rapid scaling.
Peter Smith
