Lessons We Can Gather from Avocado: Meta’s Unreleased AI Model

      In the competitive arena of AI agents, where companies are finalizing investment deals daily to enhance their AI infrastructure and software, the organizations appearing to be at the forefront of the race include OpenAI, Anthropic, Microsoft, NVIDIA, Google, and Amazon. However, even with the success of its large language models (LLMs), Meta is one of the big tech firms struggling to maintain its relevance.

      Currently, Meta's AI strategy is fragmented, balancing openness, scalability, and control. The launch of its new AI model, codenamed 'Avocado', has been delayed due to performance issues, inciting discussions in the industry regarding open-source and profitability.

      **Meta’s AI Strategy: Meta AI, LlaMa, and 'Avocado'**

      Led by Mark Zuckerberg, Meta has been rapidly advancing its position in the field of artificial intelligence with Meta AI. Initially released as a chatbot for WhatsApp, Instagram, Facebook, and Messenger in September 2023, the product saw significant progress with the introduction of a standalone app in April 2025 at Meta’s LlamaCon developer conference, featuring a Discover Feed, voice capabilities, and enhanced personalization options. It was created as a consumer interface for generative AI, enabling users to create content, converse, interact via a Discover Feed, and run advertisements within Meta's ecosystem.

      Behind Meta AI lies LlaMa, the company's family of LLMs. LlaMa was initially designed to assist researchers and others lacking access to extensive infrastructure to explore AI models, positioning itself as a means to democratize access within the industry. Over time, LlaMa introduced four models as an open-source multimodal AI system. Moreover, Meta provided a limited preview of the LlaMa API (Application Programming Interface) to allow developers to connect and utilize their LlaMa models.

      However, beyond LlaMa and Meta AI, reports indicate that Meta has been developing its next generation of AI models, named 'Avocado'. Although no official announcement has been made, a Meta spokesperson conveyed to Reuters that work on this new AI model differs from their prior models. While LlaMa is known for its open-source nature, 'Avocado' is set to be proprietary, preventing external developers from freely downloading its weights and related software components.

      **Avocado Isn’t the Main Story, What it Reveals About Meta Is**

      Meta’s shift raises questions about the unique aspect that Zuckerberg took pride in back in 2024, which centered around the idea that open-source could bridge the gap in AI development by enabling developers to enhance it and create smaller variations. A year after that initial memo, Zuckerberg released another memo expressing his hope for Meta to continue as a leader in Open Source. However, he indicated that safety concerns would lead them to be cautious about “what we choose to open source,” hinting at a reconsideration of his original stance.

      This decision can be viewed as a transition for Meta from a clear AI strategy to a reactive approach. The rise of Deepseek as a formidable competitor in the AI domain, with its R1 model and a range of smaller distilled variants based on LlaMa and Qwen architectures, underscored that open-source components could be utilized to establish highly competitive systems. This put Meta at a disadvantage, as its open-source models provided significant leverage to a rival. Additionally, a closed-source AI model offers financial relief for the substantial investment Meta is making to enhance its AI capabilities. For instance, in June 2025, Meta invested $14.3 billion in data labeling company Scale AI for a 49% share, bringing Scale AI’s founder, Alexandr Wang, on board to lead the new Meta Superintelligence Labs, responsible for developing 'Avocado'.

      **Disruption: Meta Under Pressure**

      Recent developments, however, indicate that Meta’s AI strategy may be entering a phase of uncertainty. Initially positioned as a leader in open-source models with LlaMa, the launch of the latest generation faced significant challenges. The initial reception of the new model was mixed, with some developers indicating it underperformed compared to competing systems, resulting in a drop in adoption compared to earlier models.

      Additionally, the release of the flagship Llama 4 model, ‘Behemoth’, which was anticipated to be a substantially larger “teacher model,” has been repeatedly delayed as engineers work to enhance its capabilities. The new frontier model ‘Avocado’ was expected to debut in March 2026 but is also facing challenges. A source familiar with the situation informed Reuters that its release has been postponed to May or June because 'Avocado' did not meet expectations against Google’s Gemini 2.5, Gemini 3, and other models from competitors in internal tests for reasoning, coding, and writing.

      Moreover, insiders revealed that Meta's leadership is reportedly considering temporarily licensing Gemini from Google to support 'Avocado' and its other AI products, although no decisions have been finalized.

      **Implications**

      The sequence of these events suggests