What we can learn from Avocado: Meta's unreleased AI model.

      In the competitive realm of AI agents, where daily investment deals are being made to develop and enhance AI infrastructure and software, the front-runners appear to be OpenAI, Anthropic, Microsoft, NVIDIA, Google, and Amazon. However, amidst the success of its family of large language models (LLMs), Meta, one of the major tech giants, seems to be struggling to maintain its relevance.

      Currently, Meta's AI strategy is divided among openness, scale, and control. Its anticipated launch of the new AI model, codenamed ‘Avocado’, has been delayed due to performance issues, prompting discussions in the industry regarding open-source versus profitability.

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

      Under Mark Zuckerberg's leadership, the company has been ramping up its focus on artificial intelligence with the launch of Meta AI. Initially introduced as a chatbot integrated into WhatsApp, Instagram, Facebook, and Messenger in September 2023, it made significant progress in April 2025 with the release of a dedicated standalone app at Meta’s LlamaCon developer conference, which included features like a Discover Feed, voice capabilities, and improved personalization.

      This app was designed as a user-friendly interface for generative AI, allowing users to create content, engage in conversations, explore through the Discover Feed, and run ads within Meta's ecosystem. Meta AI is underpinned by LlaMa, Meta’s collection of LLMs. LlaMa was first released as a tool to assist researchers and others with limited access to substantial infrastructure for studying AI models, aiming to democratize access within the industry.

      Since its inception, LlaMa has introduced four models as an open-source multimodal AI system. Moreover, Meta has provided a limited preview of the LlaMa API (Application Programming Interface) to enable developers to utilize their LlaMa models.

      Beyond LlaMa and Meta AI, reports indicate that Meta is also developing its next generation of AI models, ‘Avocado’. While there hasn't been an official announcement, a Meta spokesperson clarified to Reuters that the company is working on this new AI model, which will differ from their past models. Unlike LlaMa’s open-source framework, ‘Avocado’ will be proprietary, restricting external developers from freely downloading its weights and associated software components.

      The implications of Avocado extend beyond its existence. Meta’s shift reflects a challenge to the differentiation strategy Zuckerberg was proud to uphold in 2024, which centered on the premise that open-source would bridge the gaps in AI development by allowing developers to enhance and create smaller versions. A year after that initial communication, Zuckerberg issued another memo expressing a desire for Meta to remain a leader in open-source while indicating that, due to safety concerns, the company would be more judicious about what to make open-source, hinting at a reconsideration of his earlier stance.

      This change can be seen as a transition in Meta's approach from a distinct AI strategy to a more reactive one. Following the emergence of Deepseek as a significant competitor with its R1 model and various smaller distilled versions built on LlaMa and Qwen architectures, it became evident that open-source components could be successfully utilized to create highly competitive systems, placing Meta at a disadvantage. Closed-source models could potentially ease the financial burden stemming from Meta's substantial investments aimed at enhancing their AI capabilities.

      For example, in June 2025, Meta invested $14.3 billion in data-labeling company Scale AI for a 49% stake, appointing founder Alexandr Wang to spearhead the newly established Meta Superintelligence Labs, now responsible for developing ‘Avocado’.

      Disruption: Meta Under Pressure

      Recent developments suggest that Meta's AI strategy is entering a phase of uncertainty. While initially positioning itself as a leader in open-source models with LlaMa, the latest generation's rollout has encountered significant challenges. The initial response to the new model was lukewarm, with developers reporting underperformance compared to competitors, alongside a decline in engagement and adoption compared to previous models.

      Furthermore, the anticipated release of Llama 4’s flagship model, ‘Behemoth’, which was expected to be a more advanced "teacher model," has faced repeated delays as engineers continue to work on enhancing its capabilities. The launch of ‘Avocado’, originally slated for March 2026, appears similarly troubled, with sources indicating it may be postponed to May or June due to its underperformance relative to Google’s Gemini 2.5 and Gemini 3, as well as other AI competitors in internal assessments of reasoning, coding, and writing capabilities.

      There have also been indications that Meta's leadership is contemplating temporarily licensing Gemini from Google to support ‘Avocado’ and other AI products, though no decisions have been finalized.

      Implications

      Collectively, these developments raise questions about Meta’s long-term AI strategy. It seems that Meta is no longer implementing a cohesive AI strategy but rather exploring multiple avenues simultaneously. The core of