Insights We Can Gain from Avocado: Meta's Unreleased AI Model

      In the competitive world of AI agents, where companies are making investment deals daily to enhance their AI infrastructure and software, the frontrunners appear to be OpenAI, Anthropic, Microsoft, NVIDIA, Google, and Amazon. However, despite its success with large language models (LLMs), one major tech player struggling to maintain relevance is Meta.

      Meta's AI strategy is currently divided among openness, scale, and control. The launch of its new AI model, code-named 'Avocado,' has been delayed due to performance issues, sparking a discussion within the industry regarding open-source versus profitability.

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

      Under Mark Zuckerberg's leadership, Meta has been actively positioning itself in the AI landscape with the introduction of Meta AI. Initially launched as a chatbot on WhatsApp, Instagram, Facebook, and Messenger in September 2023, it made significant strides in April 2025 with a standalone app revealed at Meta’s LlamaCon developer conference. This app features a Discover Feed, voice capabilities, and enhanced personalization, serving as a consumer-facing interface for generative AI, enabling users to generate content, converse, and interact with ads within Meta's ecosystem.

      Internally, Meta AI relies on LlaMa, its family of LLMs, originally created to provide researchers without significant infrastructure access to AI models, thus democratizing access in the field. LlaMa has since released four models as an open-source multimodal AI system, and Meta has introduced a limited preview of the Llama API (Application Programming Interface) for developers to utilize their models.

      However, reports indicate that Meta is developing a new generation of AI models named 'Avocado.' While no official statement exists, a Meta spokesperson mentioned to Reuters that this new model differs from previous iterations. Unlike Llama's open-source approach, 'Avocado' would be proprietary, preventing outside developers from freely accessing its weights and related software components.

      **The Implications of Avocado**

      Meta's shift in strategy challenges the differentiation it once celebrated in 2024, following Llama's deployment. Initially, the open-source model was believed to bridge the gap in AI development, allowing developers to enhance it and produce smaller versions. A year later, Zuckerberg noted a need for caution regarding what to make open source due to safety concerns, indicating a reconsideration of his initial stance.

      This change can be viewed as a transition from a proactive AI strategy to a more reactive one, particularly after the rise of Deepseek, which constructed competitive systems using open-source components based on Llama and Qwen architectures. This provided Deepseek with a significant advantage, as its open-source models hampered Meta's competitiveness. Moreover, a closed-source model could alleviate some economic pressure from the substantial investments Meta is making in AI technology.

      For example, in June 2025, Meta invested $14.3 billion in data-labeling company Scale AI, acquiring a 49% stake and bringing Scale AI's founder, Alexandr Wang, into the newly established Meta Superintelligence Labs, responsible for developing 'Avocado.'

      **Meta Under Pressure**

      Contrary to its aspirations, recent events indicate that Meta’s AI strategy may be facing uncertainty. Initially seen as a leader in open-source models with Llama, the latest model rollout encountered notable obstacles. The early feedback indicated mixed results, with some developers pointing out underperformance relative to competitors, leading to lower adoption rates compared to earlier models.

      Additionally, Llama 4's flagship model, ‘Behemoth,’ has faced repeated delays as engineers continue to work on improvements. The highly anticipated 'Avocado' was initially set to launch in March 2026 but has also been delayed, now projected for May or June, reportedly falling short against Google’s Gemini 2.5 and 3 models in reasoning, coding, and writing capabilities during internal evaluations. Furthermore, discussions within Meta's leadership have emerged about temporarily licensing Google's Gemini to enhance 'Avocado' and other AI products, though no decisions have been finalized.

      **Final Reflections**

      These developments suggest that Meta may no longer be pursuing a singular, cohesive AI strategy but is instead exploring multiple pathways. The core of this shift lies in the increasing conflict between openness and control. Llama's triumph positioned Meta as a key participant in the open-source AI ecosystem, facilitating widespread adoption and growth beyond the company’s confines.

      The challenge has become maintaining a competitive advantage, particularly as rivals like DeepSeek leverage open-source structures to advance their models. The introduction of 'Avocado' indicates a strategic pivot from open to closed models, which may undermine Meta's differentiation as an open AI provider for all.

      Moreover, this alteration offers a potential solution to offset the costs incurred from Zuckerberg's heavy investment in AI development, illustrated by a commitment of $600 billion towards U.S. AI infrastructure, data centers, energy projects, and workforce initiatives by 2028. These expenditures are critical for the company